AImnesia is a persistent memory system for Claude, GPT, and local LLMs. Runs on your own machine via Docker. BM25 + vector search with nightly refinement, bi-temporal schema, and a Model Context Protocol server so Claude Desktop remembers between conversations — without surrendering your data to anyone's cloud.
Context windows reset. ChatGPT "memory" stays inside OpenAI. Claude Projects don't talk to GPT. Local models forget everything on restart. You've explained your project twenty times.
OpenAI's memory stays with OpenAI. Anthropic's Projects stay with Anthropic. Nothing crosses providers.
Your private notes, research, and conversation history — processed and stored on someone else's servers.
Flat vector stores return fuzzy matches. No temporal awareness. No conflict detection. No consolidation.
Postgres + pgvector under the hood. Eight features that turn raw transcripts into searchable, self-organizing memory.
BM25 full-text + vector similarity, merged via Reciprocal Rank Fusion and reranked with a cross-encoder. Finds things keyword search misses and things vector search misses.
RRF + cross-encoderMCP server for Claude Desktop, REST API for GPT/ChatGPT custom actions, Python client for local models. One memory layer, every assistant.
MCP · REST · PythonA refiner splits weak memories. A synthesizer clusters related chunks. A dream pass surfaces analogical bridges. Your corpus gets better while you sleep.
7 scheduled passesEvery memory tracks valid_from, valid_until, and created_at. Superseded facts are soft-expired, never destroyed. You can always replay what you knew on any date.
Auditable historyDynamic memories decay. Static (synthesized) memories persist. Anchor memories are write-protected. Restricted memories store pointers without content.
4-tier protectionTwo-way sync with an Obsidian vault — five-level hierarchy from hubs to leaves. Your memory graph becomes a navigable notebook.
Markdown-nativeNightly scans surface contradictions between memories. You review and resolve — the system doesn't silently pick a winner.
Human-in-the-loopSha256-checksummed migration runner applies pending changes on container start. Upgrade safely; never touch ALTER TABLE again.
Automatic migrationsIf you've used Docker Compose before, you're thirty seconds from a working memory layer. If you haven't, the README walks you through it.
Pull the repo, copy .env.example to .env, and drop in your Voyage AI and OpenRouter keys. That's the only config you touch.
One command brings up Postgres+pgvector, the API server, the intake server, and the nightly pipeline. Schema migrations run automatically on first boot.
Point Claude Desktop at the MCP server, hit the REST API from GPT, or query directly from Python. Your AI now has long-term memory.
Tested on Windows 10/11 (Docker Desktop), macOS (Apple Silicon & Intel), and Linux (Ubuntu 22.04+). No GPU required for the base system.
No subscription. No seat counts. No telemetry pinging home. You own the license, the data, and the machine it runs on.
AImnesia is the memory layer I built for my own assistants because nothing on the market would run on my machine. Now it can run on yours.