The recall command is the primary interface for Recall. All subcommands operate on the SQLite database at ~/.agents/Recall/recall.db.
New to Recall? Start with the Getting Started first-run tutorial.
recall "query" # Hybrid search (keyword + semantic, default)
recall "query" -k # Keyword only (FTS5)
recall "query" -v # Semantic search (requires Ollama)
recall search "query" # FTS5 search with options
recall search "query" -t decisions # Hard-filter to decisions only
recall search "query" --bias-type decisions # Prefer decisions, still show other matching tables
recall search "query" -p myproject # Filter by project
recall search "query" --show-provenance # Show provenance for every result
recall search "query" --include-duplicates # Include records marked by recall dedup
recall semantic "query" # Semantic search (explicit)
recall hybrid "query" # Hybrid search (explicit)The default search mode is hybrid: it combines FTS5 keyword matching with vector similarity, then merges results by relevance score. Use -k when you want exact phrase or operator matching without semantic expansion. Use -v when you want conceptually related results even if the exact words differ. -k, -v, and recall search stay local even when ~/.agents/Recall/config.json names a query harness. recall hybrid and an unqualified recall "query" use that harness when the query key is present. A missing key stays hybrid. See architecture.
Local semantic and hybrid search skip stored vectors whose length differs from the query embedding and print one warning with the skipped count. Compatible vectors still contribute to ranking. If every stored vector is incompatible, semantic search falls back to keyword search, while hybrid search returns its keyword matches. The warning recommends recall embed rebackfill to regenerate incompatible embeddings.
For keyword search, -t/--table and --bias-type solve different problems:
-t decisionsis a hard filter: only decisions can appear.--bias-type decisionsis a soft boost: matching decisions rank higher, but matching LoA entries, learnings, breadcrumbs, and messages can still appear.
Allowed values: messages, loa, decisions, learnings, breadcrumbs.
recall search "database choice" -t decisions # only decisions
recall search "database choice" --bias-type decisions # decisions first, broader context preserved
recall "database choice" -k --bias-type decisions # default command, keyword-only with soft biasFTS5 supports boolean operators and prefix matching:
auth AND token— both terms must appearconfig OR setup— either termconfig NOT docker— exclude a termauth*— prefix match (authz, authentication, etc.)"vpn config"— exact phrase
By default, search output stays quiet about Record Provenance when a record carries a known value, and visibly flags records whose provenance is unknown (legacy rows that predate the provenance column). Pass --show-provenance to display the provenance of every result.
Records marked as duplicates by recall dedup are hidden from every search path (keyword, semantic, hybrid) by default — the records and their lineage remain in the database. Pass --include-duplicates to show them.
Public door for harness adapters. Send raw text; core ranks and stores. No SDK. Guide: Capture adapter. Hidden recall host-hook is deprecated for ambient writes. It remains only for Codex SessionStart injection.
recall capture --contract 1 --harness acme-agent --event turn_end --text "raw turn"
recall capture --contract 1 --harness omp --event session_end --text-file turn.txt
printf '%s\n' '{"contract":1,"harness":"acme-agent","event":"turn_end","text":"raw turn"}' | recall capture--contract defaults to 1. --harness is ^[a-z][a-z0-9-]{0,63}$; see built-in and reserved harness IDs. --event is turn_end or session_end. Optional: --session-id, --cwd, --project. Text is --text, --text-file, or stdin when those flags are absent and another flag selected the flag form. With no flag fields, stdin is the JSON twin (contract, harness, event, text, session_id, cwd, project). Invalid input exits non-zero and writes nothing.
Prove a write with recall search "<unique phrase>" -t messages.
recall dump "Session Title" # Import session + capture LoA entryCombines recall import and Curated LoA capture in one step. Run this at the end of every session to persist the conversation and run the fabric Extractor. The title should describe what was accomplished in the session. --skip-fabric writes a basic summary instead.
Without --stdin, Recall discovers native transcripts in Claude Code, OpenCode, then Pi order, followed by registered session sources. For omp's live-session route and discovery restriction, see the recall_dump contract.
--stdin reads up to 25 MiB of contract 1 JSON for host extensions that hold the live transcript: contract: 1, harness, session_id, optional cwd/project, and messages with user/assistant role, nonblank text content, and optional timestamp. Malformed input is rejected before writes. -p overrides the payload's project; otherwise Recall uses its project, then the origin repository name or directory basename from cwd, then harness if cwd is absent.
Both CLI dumps and MCP memory_dump scrub each message's content with the shared write-safety guard before snapshot matching, storage, and summary or extraction. Known credential patterns become visible [REDACTED:<kind>] markers, invisible/bidi controls are removed, and incoming messages left blank are dropped. This does not detect every secret; follow the credential-handling rule before dumping sensitive conversations.
recall loa write "Session Title" # Capture messages since last LoA entry
recall loa write "VPN Config" -p infra # Tag with project
recall loa write "VPN Part 2" -c 1 # Continue from previous entry (chain)
recall loa write "Auth System" -t "auth" # Add tags
recall loa list # List recent entries
recall loa show 1 # View extract body for entry #1
recall loa quote 1 # View raw source messages for entry #1LoA is the primary knowledge capture mechanism. Raw transcripts are high-noise; Curated LoA entries contain fabric Extractor output, message lineage, continuation chains, and project context. Chaining (-c <id>) links a new entry to a prior one, forming a thread of related sessions. Automatic-capture LoA is written by session extract, not this command. Extractor config: architecture.
The omp extension invokes recall loa omp-end <sessionId> for background curation.
recall add decision "Use TypeScript over Python" --why "Type safety, team preference" -p myproject
recall add decision "Use SQLite for storage" -w "Lightweight, zero-config" --confidence highRecords an architectural or strategic decision with rationale. Decisions carry a status field (active, superseded, reverted) so you can track when a decision is overturned and why.
Manage decision status transitions.
# List all decisions (all statuses)
recall decision list
recall decision list --status active
recall decision list --status superseded
recall decision list --project my-project --limit 10
# Mark a decision as superseded (replaced by newer)
recall decision supersede 42
# Mark a decision as reverted (was wrong, rolled back)
recall decision revert 42recall add learning "Port conflict on 4000" "Kill process or change port" --prevention "Use dynamic port allocation"Captures a problem-solution pair. The first argument is the problem, the second is the solution. The optional --prevention flag records how to avoid the issue in the future.
For learning-tag scrubbing and secret-handling requirements shared with memory_add, see the explicit-write reference.
recall add breadcrumb "User prefers dark mode in all UIs" -p myproject -i 8Stores a freeform observation or preference. The -i flag sets importance on a scale of 1–10 (default: 5). Higher importance surfaces the breadcrumb earlier in search results.
recall jev "Use SQLite, not a second database" # Score one candidate
recall jev --kind decision -p recall "Use SQLite" # Optional kind and project
printf '%s\n' "hi" | recall jev # Read the candidate from stdinPrints one JSON object with choice (keep, demote, or drop), probabilities, and confidence. Does not open the database and does not write a row. A non-blank JEV_RECALL_KEY in the process wins; otherwise Recall reads it from $HOME/.env. If neither provides a key, a terminal prints https://console.typesafe.ai/keys and prompts for one run. The paste is not saved or printed. A non-interactive run prints that URL and exits 1. Ingest uses the same resolution without prompting; no resolved key skips Jev and still writes parsed rows. Setup is in Score one memory item with Jev.
recall loa list # Recent LoA entries
recall loa show 1 # Extract body for entry #1
recall loa quote 1 # Raw source messages for entry #1
recall recent # Recent records across all tables
recall recent decisions # Recent decisions only
recall show decisions 5 # Full record for decision #5
recall stats # Database statistics
recall doctor # Health check all subsystemsrecall recent accepts any table name: decisions, learnings, breadcrumbs, loa, sessions, docs, telos.
recall show displays the full record for a single row. The first argument is the table name, the second is the row ID.
recall import --dry-run # Preview session imports without writing
recall import --yes -v # Import all sessions from ~/.claude/projects/
recall import-legacy --yes # Import DISTILLED.md extracts as LoA entries
recall docs import --dry-run # Preview document imports
recall docs import --yes # Import standalone markdown documents
recall telos import --dry-run # Preview TELOS imports
recall telos import --yes -u # Import TELOS framework (update existing)
recall telos list -t goal # List TELOS entries of type: goal
recall telos show G7 # Show a specific TELOS entry by IDrecall import scans ~/.claude/projects/ for JSONL session files and imports any not yet recorded. Use --dry-run first to confirm scope. The -v flag enables verbose output.
recall import-legacy converts entries from DISTILLED.md files (the previous manual memory format) into LoA entries, preserving their content in the new structured format.
recall docs manages standalone reference documents (markdown files) that should be searchable alongside session memory.
recall telos manages TELOS framework entries (goals, principles, constraints). The -u flag updates existing entries rather than skipping them on reimport.
recall embed stats # Show stored embedding counts and storage size
recall embed backfill -t loa # Generate missing embeddings for LoA entries
recall embed backfill -t decisions # Generate missing embeddings for decisions
recall embed rebackfill # Regenerate existing embeddings with the current modelEmbeddings are generated via Ollama and stored in the database for vector similarity search. recall embed stats groups stored embedding counts by source table and model and reports their total storage size. Run recall embed backfill after importing a large batch of records to generate missing embeddings.
Supported tables for backfill: loa, decisions, learnings, messages (assistant messages only). Each run processes up to --limit records (default: 100).
recall embed rebackfill regenerates existing embeddings with the current model before replacing them in one transaction, then rebuilds the vector index when available. If embedding generation fails, the stored embeddings remain unchanged. Embeddings whose source is no longer available or whose text is too short are removed during replacement.
recall start # Render L0/L1 session-start memory (markdown)
recall start --format cursor # Cursor sessionStart wrapper: { "additional_context": "..." }
recall onboard # Interactive L0 identity interview (writes identity.md)
recall onboard --print --yes # Preview rendered markdown without writing
recall onboard --project # Write project-local (./.atlas-recall/identity.md)
recall onboard --out /path/identity.md # Write to an explicit pathrecall start is the shared L0/L1 inject renderer. Claude SessionStart and Codex
SessionStart call the same assembler (published caps: L0 1200 chars, L1 6000,
total 8000). Empty DB + missing identity.md exits 0 with a short degrade line.
Cursor inject uses --format cursor ({ additional_context }); Codex keeps
hookSpecificOutput.additionalContext. The templates/cursor/ sessionStart
command stays unqualified recall start --format cursor. Cursor.app GUI PATH
typically lacks ~/.bun/bin, so that hook is a no-op until recall is on the
app PATH. CLI Cursor / a shell where recall resolves is fine. Durable GUI
PATH / recall start --format cursor accuracy is pending FM-321/327; this
section describes the as-built unqualified command, not a wrapper or
machine-specific prefix.
recall onboard creates the L0 tier that recall start injects at the top of every
session. Precedence for the output path: --out > RECALL_IDENTITY_PATH env var >
--project > the global identity. The global resolver preserves an existing
user-owned ~/.claude/MEMORY/identity.md; otherwise it selects the canonical
file under the Recall install root, ~/.agents/Recall/MEMORY/identity.md. If a
file already exists, the command asks for confirmation and writes a .bak copy
before overwriting.
The renderer warns when output exceeds MAX_L0_CHARS=1200 — recall start
silently truncates beyond that threshold.
The importance column (1-10) on messages, decisions, learnings, and
loa_entries controls L1 tier ranking at session start.
# Backfill importance scores from confidence signals (dry-run by default)
recall importance backfill
recall importance backfill --execute
recall importance backfill --execute --table decisions
recall importance backfill --execute --force # overwrite non-defaults too
# Pin / unpin individual records
recall pin decisions 42 # Pin decision #42 to importance 10 (default)
recall pin learnings 7 8 # Pin learning #7 to importance 8
recall unpin decisions 42 # Reset to table default (5, or 8 for LoA)LoA entries have a write-time floor of 5; recall pin will not drop them below that.
The provenance column on messages, decisions, learnings, breadcrumbs,
and loa_entries declares how each record was created: verbatim (exact source
text), user_authored (directly authored via a user or agent command),
extracted (generated from source material, possibly lossy), or derived
(mechanically produced from existing memory records). Provenance is automatic
write-path metadata — every write path stamps it; there is no flag or MCP
parameter to set it (see docs/adr/0001-record-provenance-automatic-write-path-metadata.md).
Legacy rows that predate the column have no declared provenance (NULL,
reported as unknown). The backfill classifies them conservatively — only
where the source table or a write-path marker gives deterministic evidence;
everything else stays unknown rather than being guessed:
recall provenance backfill # Dry-run report (default)
recall provenance backfill --execute # Apply the classification
recall provenance backfill --execute -t loa_entries # Limit to one tableAllowed -t/--table values: messages, decisions, learnings,
breadcrumbs, loa_entries, all (default).
Phase 2 benchmark harness for measuring context efficiency.
recall benchmark list # Show available suites + build status
recall benchmark run # Run all available suites
recall benchmark run B # Run a specific suite (A-E)
recall benchmark run B -p my-project # Scope to a specific project
recall benchmark report # Show the latest reportSuite B (token efficiency) compares the v2 wake-up bundle against v1 and the
CLAUDE.md baseline. Results are written to docs/benchmarks/results/ as JSONL plus
a human-readable .md alongside. See docs/benchmarks/README.md for methodology.
Portable and disaster-recovery exports of the memory database.
recall export # JSON export to stdout (summary on stderr)
recall export --format markdown # Human-readable Markdown export
recall export --format sql --output dump.sql # Textual SQL dump (schema + INSERTs)
recall export --format sqlite --output copy.db # Full database backup (VACUUM INTO)
recall export --output exports/ # Directory mode: artifact + manifest.json
recall export --backup # Timestamped SQL dump to ~/.agents/Recall/backups/Formats:
- json / markdown — app-level export of the durable memory tables
(
sessions,messages,decisions,learnings,breadcrumbs,loa_entries,dedup_lineage,host_ingest_generations,host_ingest_generation_messages,host_ingest_embedding_invalidations,host_ingest_state,host_ingest_messages,loa_message_sources). Every row of a provenance-bearing table carries an explicitprovenancefield; legacyNULLprovenance is exported as the literalunknown— never omitted, never guessed (see Record Provenance above). Embeddings are excluded. - sql — textual SQL dump (CREATE TABLE + INSERT statements) of the same durable tables. Restorable into an empty database; one-command restore is intentionally not provided.
- sqlite — binary database backup via
VACUUM INTO: the full DB including embeddings and internal tables. Always requires--output.
Output contract:
- No
--output: export data goes to stdout, the manifest summary to stderr — stdout stays clean for piping. --output <file>: writes a single export file and prints the manifest summary to stdout.--output <dir>: writes the export artifact plusmanifest.jsoninto the directory. Directory exports always writemanifest.json. A path that does not exist yet is treated as a file — add a trailing slash (exports/) to request a new directory.
The manifest records the Recall version, timestamp, schema version
(PRAGMA user_version), format, included tables, row counts per table,
provenance counts per table (including unknown), and whether embeddings were
included.
recall export --backup writes a timestamped SQL dump to
~/.agents/Recall/backups/ (creating the directory if needed), never
overwrites an existing file (a -N suffix is added on collision), and prints
the output path.
Detect and mark duplicate memory records without erasing evidence or lineage.
recall dedup # Dry-run report (default — writes nothing)
recall dedup --execute # Mark duplicates (non-destructive)
recall dedup --execute --delete # Destructive opt-in: hard-delete duplicates
recall dedup -t breadcrumbs # Scope to one table
recall dedup -p myproject # Scope to one project
recall dedup --threshold 0.98 # Stricter semantic matching (default 0.95)
recall dedup --no-semantic # Exact/normalized text pass onlySafety model:
- Dry-run by default. Mutations require
--execute. - Non-destructive by default.
--executewrites lineage rows to thededup_lineagetable (survivor, duplicate, reason, similarity, status, timestamp); the duplicate records themselves stay intact and are merely hidden from search.--deleteis the destructive opt-in and requires--execute— runrecall export --backupfirst. - Within-table only. Dedup never merges across tables (or across projects). Cross-table duplicate candidates are report-only.
- Survivor priority is
user_authored > verbatim > extracted > derived > unknown(Record Provenance); ties break by richness (longer normalized text), importance, recency, then lowest id. - Detection combines exact/normalized text matching with semantic
matching over stored embeddings (no embedding service call needed). The
semantic pass is skipped — and reported as skipped — when no embeddings
exist; records are never merged below the configured
--threshold(conservative default: 0.95 cosine similarity). Records with fewer than 20 significant characters are never candidates. - Survivors are sticky. A record recorded as a survivor in
dedup_lineageis never re-marked as a duplicate by a later run, so every hidden or deleted record keeps a visible survivor across runs. The tradeoff is order-dependence: an early survivor is never consolidated under a later, higher-priority record. As defense-in-depth,--deleterefuses outright (no changes written) if a plan would delete a recorded survivor. A consequence of stickiness is an intentional asymmetry: because a sticky survivor is excluded from the duplicate side of later scans, it never acquires new semantic duplicates after the run that established it. The exact/normalized pass is unaffected — exact-duplicate chains share normalized text, so a survivor that gathers further exact matches stays safe by construction. - Marking is terminal until acted on. Once
--executemarks a record a duplicate, a later--execute --deleterun does not auto-escalate that prior mark to deletion —--deleteacts only on duplicates identified within its own run. Marking is the conservative, reversible state; silently escalating it to irreversible deletion would violate non-destructive-by-default (ADR-0003). A user who wants prior marks deleted must act on them explicitly. - Prune respects survivors.
recall prunenever deletes a record that is a recorded survivor in activededup_lineage(statusmarkedordeleted) — doing so would orphan the duplicates marked under it (they would stay hidden from search while their visible representative is gone). Withheld rows are reported in the prune summary, never silently skipped — the same conservative stance as--delete's survivor guard above. - Lifecycle-aware. Only
activedecisions participate; superseded and reverted decisions are managed by the decision lifecycle, not dedup.
Marked duplicates are hidden from all search paths by default; see
recall search --include-duplicates. Lineage is included in
recall export, so the audit trail is portable.
Explicit data/index maintenance is deliberately separate from
recall doctor --fix, which repairs install and configuration drift but never
repairs memory data or indexes.
recall repair # Dry-run report (default — writes nothing)
recall repair --execute # Apply the planned repairs
recall repair -t messages # Scope to one table
recall repair --no-embed # Skip the embedding passWhat repair covers:
- FTS5 index rebuild. Each search index (
messages,decisions,learnings,breadcrumbs,loa_entries,telos,documents) is checked against its source table — indexed row count (via the docsize shadow table), sync-trigger presence, and the FTS5integrity-checkcommand. A drifted index is rebuilt from its source table; a missing index or missing sync triggers (the classic symptom: search silently returns nothing on an un-migrated database) are recreated from the canonical schema DDL and then rebuilt. - Lifecycle message publication. Pending lifecycle generations are checked
separately because their searchable rows are activated by a publication
pointer. With
--execute, repair advances their FTS publication and semantic invalidations in bounded batches. If work remains after that bounded pass, repair exits nonzero so it can be retried safely. - Re-embedding. Rows expected to carry embeddings (
loa_entries,decisions,learnings, assistantmessages) that have none are re-embedded when the Ollama embedding service is available and the row has enough source text. If the service is unavailable, repair reports the missing embeddings and still exits successfully — unless another requested repair failed. Partial results are never hidden: embedded, skipped (too short), and failed counts are all reported. - Orphan/invariant recovery. With
--execute, embeddings whose published source record no longer exists are removed and the vector index is rebuilt when sqlite-vec is available (otherwise it remains durably marked for the next rebuild). Dedup lineage pointing at missing rows, messages without a session, broken LoA message ranges and parent links, and pending schema migrations remain report-only. Pending migrations are fixed byrecall init.
Safety model:
- Dry-run by default. Mutations require
--execute. Runrecall export --backupbefore applying repairs. - Repair never hard-deletes source records. Only derived orphan embedding rows are removed.
- Repair never changes Record Provenance. FTS
rebuild regenerates index shadow tables; re-embedding only inserts into
the
embeddingstable. No source-table column is written. recall doctormay recommendrecall repair, butdoctor --fixnever runs data repair implicitly.
recall init # Initialize the database (safe to re-run)
recall doctor # Health check all subsystems
recall doctor --fix # Repair install and Claude MCP configuration drift
recall stats # Database statistics
recall path # Print resolved paths (DB, install root, symlinks)
recall path --json # Same, as JSON
recall migrate --to /new/path/recall.db # Relocate the database (details below)
recall migrate --to ... --dry-run # Preview the migration plan
recall onboard # Interactive L0 identity interview (see Onboard)
recall start # Render L0/L1 session-start memory (see Identity & Onboarding)recall init creates the database schema if it does not exist, and applies any pending migrations. It is safe to run on an existing database.
recall init and recall migrate preserve literal .. segments and repeated spaces in database paths. The filesystem resolves these paths, including link/.. through a symlink's target, when creating directories or opening and moving files.
recall doctor checks the database connection, schema integrity, FTS5 index health, MCP server registration, Ollama availability, and the per-platform symlinks under ~/.agents/Recall/. Run this first when troubleshooting. Pass --fix to repair drift: missing symlinks are re-created, stale recall-memory database paths are updated, and user-modified files at symlink targets are backed up under ~/.agents/Recall/backups/<TIMESTAMP>/doctor-fix/ before being replaced. Data and index maintenance is the explicit job of recall repair, which doctor recommends when an FTS index is out of sync.
Claude configuration files accept JSONC, but when recall install, recall update, recall uninstall, or recall doctor --fix writes one back, it is serialized as strict JSON and loses comments and trailing commas. Adding the first Recall MCP or hook entry during install or update also normalizes the file. recall migrate applies the same normalization to every detected host configuration it rewrites.
Recall's direct config editors reject malformed configuration, duplicate keys, and invalid Recall entry shapes instead of partially interpreting or silently skipping them.
recall stats reports row counts per table and total database size.
recall path prints the resolved DB path, the install root, the active env var (RECALL_DB_PATH / MEM_DB_PATH / default), and the per-platform symlink targets with their current state (OK / drift / missing). Pass --json for machine-readable output.
recall migrate moves the database to a new path and rewrites MCP configs across all detected platforms (~/.claude.json, ~/.claude/settings.json, ~/.config/opencode/opencode.json, ~/.pi/agent/mcp.json) so the spawned recall-mcp process keeps reading from the right file. When the quarterly aging wrapper exists, migration also rewrites its exported database assignment through the same staged-write and rollback path. It does not create a wrapper or change the crontab.
It expands ~ and the ~/ prefix in --to and makes relative source and destination paths absolute by prefixing the current working directory. Rewritten MCP configs retain that destination spelling. If source and destination have the same filename and their existing parent directories resolve to the same physical directory, migration exits successfully without snapshots, file moves, or config changes.
For a different destination, migration refuses to overwrite a non-empty file or proceed when any destination -wal, -shm, or -journal sidecar already exists. If lsof detects an open source database, migration also refuses to proceed; this check is best-effort when lsof is unavailable. It snapshots the source DB + WAL/SHM sidecars + configs, including their original comments, to ~/.agents/Recall/backups/<TIMESTAMP>/pre-migrate/ before any mutation, then moves the database and those sidecars. If an existing target config or aging wrapper cannot be read, safely parsed, staged for an atomic write, or replaced, migration stops before moving the database. If a later database sidecar move or staged commit fails, migration attempts to move the database and any moved sidecars back and restore the bytes and permissions of files already committed. Restart Claude Code / OpenCode / Pi after running so their MCP servers reload.
recall onboard # Interactive 7-question identity interview
recall onboard --yes # Non-interactive (accept all defaults)
recall onboard --dry-run # Show the proposed identity.md, write nothingrecall onboard runs a short interview that writes the same resolved identity path RecallStart reads — the L0 tier of tiered RecallStart. L0 is the always-loaded slice on hosts with supported session-start injection: your role, projects, tools, and working preferences. Without it, those sessions have to re-learn the basics from search.
Run it once after installing. Re-run it whenever your role, active projects, or working preferences change. The path can be overridden with RECALL_IDENTITY_PATH — honored by both recall onboard (write) and the RecallStart hook (read).
Inputs are separated with | (not , or ;) to avoid silent data loss when a value itself contains a comma: e.g. no force-push, ever | always use worktrees.
Clean up old and expired records. Dry-run by default — must pass --execute to actually delete.
# Preview what would be pruned (safe, no deletions)
recall prune
# Actually prune with default 180-day retention
recall prune --execute
# Custom retention period
recall prune --execute --older-than 90d
# Skip decision pruning
recall prune --execute --keep-decisionsWhat gets pruned:
| Table | Strategy |
|---|---|
| messages | Consolidated sessions (with LoA) older than N days |
| sessions | Orphaned (no messages, no LoA) older than N days |
| breadcrumbs | Expired (expires_at in the past) |
| decisions | Superseded/reverted older than 90 days |
| extraction_tracker | Older than N days |
Never pruned: loa_entries, learnings, extraction_errors
Dedup-aware. Prune never deletes a record that is a recorded survivor in
active dedup_lineage (status marked/deleted), even when it otherwise
matches the retention rule above — deleting a survivor would orphan the
duplicates marked under it. Withheld rows are counted and reported in the
summary (N kept as dedup survivors). See the dedup safety model.
Prune also fails closed before mutation when the lifecycle schema is not ready.
If it reports RETRYABLE, run recall init and retry the same command.
Type-aware aging. Dry-run by default. Pass --execute to apply.
recall age
recall age --executeWith the defaults, aging expires old breadcrumbs, demotes old decisions and learnings below importance 5, and deletes messages older than 180 days only from sessions consolidated into LoA entries. Existing guards still apply: recorded dedup survivors, marked duplicates, and foreign-key references are withheld. LoA entries are not demoted at the default threshold, and pinned importance 10 is never touched.
For managed cron scheduling, opt-out, paths, and lifecycle behavior, see Quarterly aging.