longmemeval

An open-source, vector-free long-term memory engine for AI agents, achieving SOTA on LoCoMo and LongMemEval with significantly less context.

Local-first AI memory — runs offline on any machine with 8 GB+ RAM (SBC, mini PC, laptop, workstation). Zero-loss verbatim archive, knowledge graph, hybrid retrieval. Framework-agnostic, no cloud.

A Multi Agent Memory MCP That Connect Agents Across Systems and Machines

Open plain-text file format for AI memory. Your assistant's long-term memory as .dai files on your disk: readable by Claude, GPT, Gemini, Cursor, local models and grep (all LLM models work). MCP server + hooks for Claude Code, Claude Desktop, Cursor, Windsurf, Codex. 83% LongMemEval-S (GPT-4o), 92% (Claude Fable 5), 10x fewer tokens.

The Cost of Remembering: filesystem memory matches long-context accuracy on LongMemEval while reading 97% fewer tokens and costing 95% less. Harness, run data, 129 agent-built memories, and paper source.

Dynamic agent memory for LLM agents: 7 neuroscience-inspired layers (working, episodic, semantic, procedural) with FSRS spaced repetition and memory consolidation. Zero-dependency TypeScript library, Model Context Protocol (MCP) server, Vercel AI SDK middleware. Paper: arXiv 2604.23878 · reproduction packages on Zenodo.

Zero-LLM agent memory for Claude Code and AI agents: local-first BM25, dense-vector, and reciprocal-rank-fusion retrieval. Returns original passages verbatim by default. Available on PyPI as fidelis-memory. Apache-2.0.

Git-versioned, contiguity-checked memory for AI agents: verbatim transcripts, byte-offset recall, no LLM

Your AI forgets everything between sessions. This fixes that — 98%+ retrieval accuracy, 100% on LongMemEval, 99% token savings. 44 MCP tools. Fully local, zero cost.

Token-native agent memory retrieval for LLMs, without embedding APIs or vector databases.

Open evaluation harness for AI agent memory systems. Runs LoCoMo and LongMemEval against Synap, Mem0, Zep and Supermemory with pluggable provider adapters.

First-Person Agent Memory Bench. 10 Categories including fact recall, multi-hop links, temporal reasoning, fact overwrites, speaker traps, refusal, credibility, and agentic tool usage. 540K token / 60 session corpus, all in first person. Dynamic output-answer-key portion. Comprehensive report with visuals and miss breakdown.

Auditable memory layer for AI agents: zero-LLM-call local ingest (~10ms/msg, air-gapped), matches Mem0 on accuracy at ~1000x lower ingest cost, bi-temporal belief-state, MCP server. Honest LoCoMo/LongMemEval benchmarks. Open source (Apache-2.0).

Open benchmark for choosing agent memory: compare Hindsight, Mem0, and OpenViking on accuracy, answer-visible context, latency, and indexing tokens/time.

Multi-agent memory substrate for PostgreSQL — provenance-gated, vector-hybrid recall

Benchmark results, scorer, and reproducibility kit for Sibyl Memory. LongMemEval 95.6% (#2). Verify it yourself.

TMCRA Core — local-first, scope-isolated long-term memory runtime for AI agents.

Long-term memory for AI agents on Jev. Keep raw records, judge them with a fast System 1 model and answer from under 4k tokens of context.

LongMemEval 中文子集：识流基于 DeepSeek-V4-Flash 的 500 题公开评测结果与可复核数据。

Reproducible benchmarks for execution-intent memory in long-horizon AI coding agents. ID-RAG cross-corpus matrix + LongMemEval-S subset; BYO API keys.