forgetmenot 🧠

Persistent, structured, semantically searchable memory for AI agents, delivered as a local MCP server written in Go. One static binary, zero runtime dependencies, data stays on your machine.

Works with any MCP-capable agent: Claude Code, Cursor, Codex, opencode and others. Zero configuration: if no embedding service is available, the built-in lexical search keeps memory working offline.

Why

Agents forget everything between sessions. You re-explain the same context, architectural decisions get lost, your preferences are ignored. forgetmenot gives your agents long-term memory: facts, decisions, preferences, entities, project context and episodes, stored locally and found semantically.

Positioning: hygiene + trust (dedupe, provenance, conflicts, intelligent forgetting) as first-class features, not add-ons. Details in PRD.md.

Features (v0.4) ✨

• memory.remember - store a memory; automatic dedupe + conflict detection + topic labels
• memory.recall - semantic search with similarity score, project/type filters, hides superseded memories, returns source + trust
• memory.timeline - trace a topic's evolution across sessions (correlation!)
• memory.forget - delete a memory
• memory.update - change content/type/project/importance/trust/session/metadata
• memory.link - relations between memories: related, supersedes, part_of
• memory.conflicts - list open conflicts
• memory.resolve_conflict - pick the winner; the loser becomes superseded
• memory.stats - memory and project counts
• Memory types: fact, preference, decision, entity, context, episode
• Sessions: memories grouped per agent session; session start/end/list
• Topics: subject labels for cross-session correlation
• Markdown export: export-md writes human/AI-readable .md per project
• LLM auto-topics: remember -auto-topics extracts topic labels with Ollama, OpenAI or Anthropic (Claude)
• Project summarization: summarize compresses stale sessions into a context summary (memory hygiene)
• Doctor: forgetmenot doctor diagnoses DB, embeddings, hooks and active session
• Web UI: forgetmenot web serves a local browser dashboard (memories, timeline, conflicts, sessions) from the same binary
• Compact embeddings: binary float32 BLOB (legacy JSON auto-migrated)
• Trust levels + sanitization (prompt-injection defense)
• CLAUDE.md bridge: bridge export + bridge import
• Memory budget: project_context -budget N
• Local embeddings (Ollama) or remote (OpenAI-compatible)
• Offline fallback: -embed auto (default) uses Ollama when reachable and a built-in deterministic lexical embedder otherwise - memory works on machines with no Ollama, no API key, nothing
• CLI recall: forgetmenot recall "query" mirrors the memory.recall tool for scripts and non-MCP agents
• Pure-Go SQLite: single static binary, no cgo, easy cross-compile
• CLI: remember, recall, capture, session, timeline, project_context, maintain, setup, bridge, export-md, export/import, stats, list, eval
• Eval harness with recall@k (20 queries), JSON output for CI

Cross-session topic correlation 🔀

Sessions group memories, topics label them. Together they answer "how did this subject evolve over time?":

forgetmenot session start -project demo      # hooks do this automatically
forgetmenot remember -content "chose JWT" -type decision -project demo -topics auth
forgetmenot session end -project demo        # hooks do this automatically

# next session, days later:
forgetmenot remember -content "switched refresh tokens to 60m" -type decision -project demo -topics auth

forgetmenot timeline -project demo -topic auth
# - [2026-08-11] we chose JWT for sessions (session a1b2c3d4) [decision]
# - [2026-08-15] we switched refresh tokens to 60m (session e5f6a7b8) [decision]

The MCP tool memory.timeline exposes the same correlation to agents.

Automatic operation (no manual steps) 🤖

forgetmenot is designed to run on its own. The user does not execute memory commands; hooks, agent instructions and background maintenance do:

• SessionStart hook → forgetmenot project_context injects the project summary automatically
• Stop hook → forgetmenot capture saves a session summary as an episode automatically
• Agent skill (.claude/skills/forgetmenot/SKILL.md) teaches the agent to recall/remember on its own
• forgetmenot maintain (cron/daemon) applies decay automatically

One-time setup in a project:

forgetmenot setup     # writes .claude/settings.json with the hooks

CLAUDE.md bridge

The agent's native memory (CLAUDE.md) stays in sync without manual work:

forgetmenot bridge export -path CLAUDE.md -project demo   # write project context into CLAUDE.md
forgetmenot bridge import -path CLAUDE.md -project demo   # ingest facts section into memory

The export writes a managed <!-- forgetmenot:context --> section. The import reads bullets from a <!-- forgetmenot:facts --> section.

CLI

forgetmenot session start|end -project demo            # session lifecycle (hooks do this)
forgetmenot timeline -project demo -topic auth         # topic evolution across sessions
forgetmenot remember -content "chose JWT" -type decision -project demo -topics auth
forgetmenot recall "which auth did we choose" -project demo          # works offline too
forgetmenot project_context -project demo -budget 4000 # session-start context injection (used by hooks)
forgetmenot capture -project demo                      # session-end capture, reads summary from stdin (used by hooks)
forgetmenot maintain                                   # decay + future compression (cron-friendly)
forgetmenot setup                                      # write Claude Code hooks config
forgetmenot setup -mcp .mcp.json                       # write .mcp.json for ANY agent (absolute path)
forgetmenot bridge export|import -path CLAUDE.md       # CLAUDE.md sync
forgetmenot export-md -project demo                    # human/AI-readable markdown
forgetmenot web -addr 127.0.0.1:8090                  # local browser dashboard
forgetmenot summarize -project demo -llm ollama        # compress stale sessions
forgetmenot doctor                                     # diagnose setup
forgetmenot stats                                      # memory + project counts
forgetmenot list -project demo                         # list memories
forgetmenot export -project demo > mem.json            # portable backup (with embeddings)
forgetmenot import < mem.json                          # restore
forgetmenot eval [-json]                               # seed + eval against real embeddings (Ollama)

Benchmark

forgetmenot eval runs a fixed 20-query dataset and reports recall@k. The dataset and runner live in internal/eval so the benchmark is reproducible:

forgetmenot eval -embed ollama          # against local Ollama embeddings
forgetmenot eval -embed openai -embed-url https://api.openai.com/v1 -embed-api-key $KEY
forgetmenot eval -embed lexical         # offline, deterministic - no service needed
forgetmenot eval -json                  # machine-readable for CI

The dataset is verified in CI (hermetic bag-of-words embedder): recall@k = 100% (20/20) on the default dataset. The built-in lexical embedder also scores 100% (20/20) offline, so forgetmenot eval works on any machine. Real-model results vary by embedder; the same command above produces yours.

Install 🚀

Requires Go 1.26+:

go install github.com/iwanro/forgetmenot/cmd/forgetmenot@latest

Or build locally:

make build
./bin/forgetmenot -version

Automatic builds: every push to main (and every PR) runs tests and cross-compiles static binaries for Linux, macOS and Windows (amd64 + arm64). Grab the latest from the CI run's artifacts — no need to wait for a tagged release.

Embeddings

Default: auto (zero configuration). The server tries local Ollama; if it is unreachable, remember/recall transparently use the built-in lexical embedder. The moment Ollama comes up, semantic search resumes - and recall automatically re-embeds any memories written during the outage, so nothing goes stale or invisible.

forgetmenot                  # -embed auto, works with or without Ollama

Strict local (Ollama):

ollama pull nomic-embed-text
ollama serve  # default: http://localhost:11434
forgetmenot -embed ollama    # fails loudly if the endpoint is down

Remote (OpenAI-compatible):

forgetmenot -embed openai -embed-url https://api.openai.com/v1 -embed-api-key $OPENAI_API_KEY

Offline-only:

forgetmenot -embed lexical   # deterministic lexical embeddings, no network at all

LLM features (auto-topics, summarize)

The optional chat provider for topic extraction and session summarization is independent of embeddings and supports Ollama, any OpenAI-compatible endpoint and the Anthropic Messages API:

forgetmenot -llm anthropic -llm-api-key $ANTHROPIC_API_KEY              # Claude
forgetmenot -llm openai -llm-url https://api.openai.com/v1 -llm-api-key $KEY
forgetmenot -llm ollama                                                 # local, default model llama3.2

-llm-model overrides the default per provider.

Agent setup (any MCP client)

One command writes a ready-to-use .mcp.json for ANY MCP client (opencode, Cursor, Codex, Claude Code...). It uses the absolute binary path (no $PATH needed) and bakes in the same -db the hooks use, so agents and hooks share one database:

forgetmenot setup -mcp .mcp.json

Or add the server manually to ~/.claude.json, .mcp.json in your project, or your agent's MCP config:

{
  "mcpServers": {
    "forgetmenot": {
      "command": "forgetmenot",
      "args": []
    }
  }
}

No Ollama, no API key, no PATH tricks required: the default auto mode makes memory.remember and memory.recall work out of the box.

If forgetmenot is not in $PATH, use the absolute path to the binary (or run go install github.com/iwanro/forgetmenot/cmd/forgetmenot@latest; the setup -mcp command handles this automatically). forgetmenot doctor warns when the binary is not on $PATH. The database is created automatically at $XDG_DATA_HOME/forgetmenot/memory.db (default ~/.local/share/forgetmenot/memory.db). Override with -db.

Usage 💬

Once connected, your agent has the memory.* tools. Examples:

Remember that the backend is FastAPI on Python 3.12, DB Postgres 16.
→ the agent calls memory.remember

Continuing feature #42. Do you know the context?
→ the agent calls memory.recall and retrieves the relevant memories

Forget the memory about the old SMTP client.
→ the agent calls memory.forget

Development

make test    # unit tests
make build   # static binary in ./bin
make lint    # go vet

Structure:

cmd/forgetmenot/    entry point, CLI subcommands
internal/memory/    core: model, SQLite store, service (remember/recall/...)
internal/embed/     embedding providers (Ollama, OpenAI-compat, lexical fallback, auto)
internal/mcpserver/ MCP layer (memory.* tools)
internal/eval/      eval harness (recall@k)

Roadmap 🗺️

• M0 ✅: remember/recall/forget/update/stats, SQLite, embeddings
• M1 ✅: relations (link/supersedes), conflicts + resolution, CLI, eval harness
• M2 ✅: automatic operation (project_context, capture, hooks, agent skill), decay + maintain
• M3 ✅: trust levels + sanitization, CLAUDE.md bridge, memory budget, public benchmark
• M4 ✅: sessions, topics, timeline correlation, markdown export, compact embeddings
• M5 ✅: Web UI dashboard, topics in recall
• v0.3 ✅: LLM auto-topics + summarize, doctor, release automation
• v0.4 ✅ (this release): zero-config embeddings (auto mode + lexical fallback), CLI recall, offline eval
• M6: HTTP/SSE transport, plugins, telemetry

License

MIT.