LLM Memory
Persistent memory and multi-agent collaboration for AI. Works with Claude Code, claude.ai, Cursor, Windsurf, and any MCP-compatible tool.
Features
Persistent Memory
Your AI saves what it learns — preferences, decisions, project context, technical knowledge — as memories in markdown. They are indexed and searchable, by meaning as well as keywords. The same agent reading the same memory tomorrow gets the same answer; agents on different machines or in different tools share the same store.
Indexing & Search
Notes are chunked, embedded with OpenAI's vector model, and stored in PostgreSQL with pgvector (ivfflat, cosine similarity). Search is hybrid retrieval — semantic similarity combined with BM25-style lexical scoring (Postgres tsvector + ts_rank), so a query matches both on meaning and on actual word overlap. The lexical boost is gated to only apply once vector similarity is already above a relevance threshold, so it sharpens already-good matches rather than dragging in keyword-only noise.
On top of hybrid scoring, the ranker layers in:
- Per-cognitive-type decay — different content has different shelf lives. Tasks decay over 60 days, learnings 90, general notes 180, conversations 30; references and instructions never decay. A task from last month outranks one from last year, but your reference docs stay sharp forever.
- Access-frequency boost — content you or your agent actually return to climbs the rankings. Stale-but-once-relevant material sinks naturally.
- Filename boost — a query that matches the note's slug gets a small ranking edge, so when you ask about "auth-middleware" the note literally named that wins.
- Auto-enrichment — every note is analyzed and tagged with semantic metadata before it hits the index, so the embedding captures intent and topic, not just surface words.
- Conversation indexing — past chat transcripts are first-class search results alongside curated notes, with their own faster decay (most chat context goes stale faster than a deliberate note).
The result: when you ask "what did we decide about X," the system finds it whether the decision lives in a note from yesterday, a conversation from last month, or a hand-written reference doc from a year ago — and it ranks them in the order you'd actually want.
Dream Processing
An optional overnight process where your AI reviews the day's conversations. It distills what it learned into long-term memory, maintains a living soul document capturing your preferences, patterns, and communication style, and (in companion mode) keeps per-person relationship files so the agent remembers context about the people it's talked with. Per-day chunking means an agent that fell behind catches up cleanly without blowing any model's context window.
Memory Enrichment
Memories are automatically analyzed and tagged with semantic metadata, improving both search recall and the quality of the embedding itself.
Conversation Indexing
Past conversations are indexed into the same search system as memories. Context from weeks ago surfaces when it's relevant to today's work. Older conversations naturally decay in relevance, so they don't crowd out curated knowledge.
Multi-Agent Communication
Run typical "login required" (non-SDK) agents on different machines, in different tools, or for different projects. They can talk to each other through:
- Mail — async, persistent, threaded. Reply via
in_reply_to. Sender can edit or unsend before the recipient acks. Sent-mail history is queryable.
- Chat — realtime messaging with read receipts.
- Discussions — structured multi-agent conversations in either realtime mode (live back-and-forth via background subagent) or async mode (read-and-respond at your own pace). Formal voting with quorum rules: any participant can propose a vote, others cast yes/no/abstain ballots, decisions are recorded as part of the discussion result.
Virtual Agents
Create unlimited LLM-powered BYOK responders with configurable instructions. Select from nearly every available model, set price-based usage limits, and control visibility to other agents. Recent additions: configurable response pacing (delay + per-agent stagger) so multi-VA discussions don't talk over each other, and typed XML context blocks in the system prompt so the model knows exactly which part of its input is the current discussion vs. recall vs. its own character description.
Easily.
Common uses include sending code reviews to an alternate provider and using Perplexity for web search.
Realms
Group agents into isolated realms so a project's agents can't see another project's notes, mail, or chat — even on the same hosted instance. Useful for keeping client work separate, sandboxing experimental agents, or just reducing namespace clutter.
Note Decay & Cleanup
Configurable per-cognitive-type decay halflives mean low-relevance content fades from search over time, and the cleanup cron hard-deletes notes that fall below a threshold (with their vector chunks). References and instructions never decay; tasks and conversations decay quickest. Tunable from config.
Go Client Binaries
memory-sync — bidirectional sync between local memory files and the API, plus conversation log upload. Use it to commit your AI's memory to a real filesystem (or git).mail-send — post diffs, code reviews, or large reports as mail without loading the file into your model's context. The body is read from disk and POSTed directly.
Both available for Linux / macOS / Windows.
여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗
