evermemo
A tiny, universal memory engine for humans and AI agents.
One small Go binary. No external services. It works as a CLI, an HTTP API, and an MCP server, so anything can remember things: your terminal, your scripts, Claude Code, Cursor, or any other agent. Run one as a hub and every agent in your organization shares the same trusted knowledge.
┌──────────────┐ ┌─────────────┐ ┌──────────────────┐
│ You (CLI) │ │ Any app │ │ AI agents (MCP) │
│ evermemo add │ │ HTTP API │ │ Claude, Cursor │
└──────┬───────┘ └──────┬──────┘ └────────┬─────────┘
└──────────────────┼──────────────────┘
┌───────▼──────┐
│ evermemo │ single binary
│ SQLite+FTS5 │ BM25 + semantic search
└──────────────┘
Install
Prebuilt binaries (macOS, Linux, Windows, amd64/arm64): grab one from Releases.
Go:
go install github.com/Evermemoai/evermemo@latest
Docker:
docker run -v evermemo-data:/data -p 7777:7777 ghcr.io/evermemoai/evermemo:latest
From source:
git clone https://github.com/Evermemoai/evermemo.git && cd evermemo
go build -o evermemo .
For production hubs, see deploy/ for systemd and Docker Compose examples, and SECURITY.md for the hardening checklist.
CLI
evermemo add "User prefers dark mode and tabs over spaces" --tags prefs,ui
echo "Deploy runs at 6pm UTC" | evermemo add --tags ops --ttl 7d # expires in 7 days
evermemo search "deploy time"
evermemo list
evermemo get mem_a1b2c3d4e5f60718
evermemo update mem_a1b2c3d4e5f60718 "Deploy runs at 7pm UTC now"
evermemo delete mem_a1b2c3d4e5f60718
evermemo export > memories.jsonl # backup / migrate
evermemo import < memories.jsonl
Interactive UI
Run evermemo with no arguments (or evermemo ui) for a Claude Code-style terminal UI:
type anything to remember it, use /search, /list, /get, /delete, /ns to manage
memories, /help for everything, /exit to leave.
HTTP API
evermemo serve --addr :7777
curl -X POST localhost:7777/v1/memories \
-d '{"content":"Invoices are due net-30","tags":["billing"]}'
curl "localhost:7777/v1/memories?q=invoice+due"
Set EVERMEMO_API_KEY=secret to require Authorization: Bearer secret on all /v1 routes.
MCP (Claude Code, Cursor, any agent)
evermemo speaks the Model Context Protocol over stdio, exposing eight tools:
add_memory, update_memory, search_memory, list_memories, get_memory,
link_memory, verify_memory, delete_memory.
Claude Code:
claude mcp add evermemo -- /path/to/evermemo mcp
Cursor / generic MCP config:
{
"mcpServers": {
"evermemo": {
"command": "/path/to/evermemo",
"args": ["mcp"]
}
}
}
Now your agent remembers things across sessions.
Shared memory for all your agents (hub mode)
Run one evermemo as your organization's memory hub, and point every agent at it. All agents share, search, and build on the same knowledge, and every memory records which agent wrote it.
# On the hub machine: one key per agent identity
EVERMEMO_AGENT_KEYS='claude:key1,cursor:key2' evermemo serve --addr :7777
# On each agent's machine: MCP proxies to the hub instead of a local file
evermemo mcp --remote https://memory.internal:7777 --key key1 --agent claude
Anything one agent stores is instantly searchable by all the others, with
provenance ("agent": "claude") on every memory. Requests with unknown keys
are rejected. EVERMEMO_REMOTE, EVERMEMO_API_KEY, and EVERMEMO_AGENT env
vars work as flag defaults. Set EVERMEMO_RATE=120 to cap each caller at 120
requests/minute. Agents can also talk MCP straight to the hub over HTTP
(POST /mcp) without a local binary.
Semantic search (optional)
By default search is SQLite FTS5 with BM25 ranking: fast, offline, zero dependencies. Point evermemo at an embedding provider and search becomes hybrid (BM25 plus cosine similarity, fused with Reciprocal Rank Fusion), so "when do we deploy" finds "release schedule is thursdays".
# Ollama (local, free)
export EVERMEMO_EMBED_URL=http://localhost:11434
export EVERMEMO_EMBED_MODEL=nomic-embed-text # default
# ...or any OpenAI-compatible API
export EVERMEMO_EMBED_URL=https://api.openai.com
export EVERMEMO_EMBED_API_KEY=sk-...
Memories are embedded on write; if the provider is down, search silently falls back to keyword-only.
For a complete, step-by-step fully-offline setup with Ollama (semantic search, consolidation, and the auto-recall proxy — all local, no cloud), see docs/ollama.md.
Trusted knowledge: provenance, confidence, graphs, ACLs
Every memory records who wrote it. On top of that:
- Verification: agents confirm or dispute each other's memories
(
verify_memorytool,POST /v1/memories/{id}/verify). Votes move a confidence score (starts at 0.6; +0.10 per confirm, -0.15 per dispute). - Memory graphs: link memories with
supersedes,relates_to, orderived_from(link_memorytool,evermemo link <from> <rel> <to>) to trace how knowledge evolved. Links come back onGET /v1/memories/{id}. - Namespace ACLs: restrict which agents can read or write which namespaces:
EVERMEMO_ACL='finbot:finance:rw,hrbot:hr:rw,hrbot:finance:r,auditor:*:r' \
EVERMEMO_AGENT_KEYS='finbot:key1,hrbot:key2,auditor:key3' \
evermemo serve
Enforced on both the REST API and the /mcp transport. If no ACL is set,
access is open.
Production hub: TLS, key rotation, backups
# HTTPS (or terminate TLS in Caddy/nginx in front)
evermemo serve --cert cert.pem --key key.pem
# Hot-reloading keys file: add/rotate/revoke agent keys without restart
cat > keys.txt <<EOF
# agent:key, one per line
claude:key1
cursor:key2
EOF
evermemo serve --keys-file keys.txt # edits picked up automatically
# Consistent online snapshot (safe while serving; uses SQLite VACUUM INTO)
evermemo backup /backups/evermemo-$(date +%F).db
Note: bearer keys travel in cleartext over plain HTTP. Always use TLS (built-in or a reverse proxy) when the hub is reachable beyond localhost.
Memory consolidation (LLM-powered hygiene)
Over time memories accumulate duplicates and contradictions. Point evermemo at a chat LLM and let it clean up:
export EVERMEMO_LLM_URL=http://localhost:11434 # Ollama; or any OpenAI-compatible API
evermemo consolidate --ns default --dry-run # see the plan
evermemo consolidate --ns default # apply it
The LLM merges duplicates, resolves contradictions (newest wins), and
archives stale memories. Nothing is deleted: sources are archived (hidden
from search, kept for audit) and linked to their replacement with
derived_from/supersedes.
Auto-recall proxy (memory without tools)
Put evermemo between your app and the LLM API, and relevant memories are
injected into every chat request automatically. No search_memory calls
needed:
evermemo proxy --target https://api.openai.com --addr :8788
# then point your SDK at http://localhost:8788 instead of api.openai.com
Works with OpenAI-style (/v1/chat/completions) and Anthropic-style
(/v1/messages) APIs, streams SSE responses through, and passes all other
routes untouched. Use --remote https://your-hub:7777 to recall from the
shared hub.
Configuration
Every command also accepts --db to point at a specific database, and --ns/namespace to partition memories per project, per user, or per agent.
Why
- Small: one binary, one SQLite file, zero dependencies to run.
- Universal: CLI for humans, HTTP for any language, MCP for any agent.
- Fast: SQLite FTS5 with BM25 ranking; millisecond search on millions of rows.
- Yours: local-first, no cloud, no telemetry.
scpthe file to back it up.
Roadmap
- Semantic (vector) search via optional embedding providers
- Memory expiry / TTL
- Import/export (JSONL)
- Streamable HTTP MCP transport
- Web dashboard (browse, search, audit, graphs)
- Webhooks / change subscriptions (reactive memory)
- Local-first replicas with hub sync
Contributing
Contributions welcome, see CONTRIBUTING.md. Found a security issue? Please follow SECURITY.md.