Mnémlet

Brain-inspired memory for your self-hosted AI stack

Runs on a Pi. No cloud. API-key protected when configured. Your data stays on your hardware.

terminal — mnemlet serve

$ mnemlet serve
╭──────────────────────────────────────────╮
│  Mnémlet v0.3 Trust                    │
│  Storage  → ~/.mnemlet                   │
│  API      → http://127.0.0.1:4050        │
│  MCP      → http://127.0.0.1:4050/mcp    │
│  LLM      → disabled                     │
│  Search   → disabled                     │
╰──────────────────────────────────────────╯
$ █

New in v0.3

Trust / Security / Privacy

API-key auth, Secret Guard, sanitized Audit logs, namespace trust policies, and backup/restore make the local release surface safer to share and demo.

Forgets Intelligently

Exponential decay + interaction-weighting. What you use stays sharp. What you ignore fades. No manual cleanup.

Thinks While You Sleep

Night consolidation: dedup, rescore, cluster, morning briefing. All local, zero API costs. Like REM sleep for your agent.

Explains Its Memory

Context packs, provenance, Audit trails, and Trust blocks show why a memory exists and whether it is active, superseded, forgotten, or confirmed.

Inspectable Vault

Every memory as a Markdown file. cat ~/.mnemlet/vault/**/*.md. Open in Obsidian. No black box.

See It Work

A v0.3 flow on an isolated demo server: authenticated remember → Secret Guard → Audit → explainable Trust.

~/mnemlet — bash

$ curl -X POST http://127.0.0.1:14060/api/v1/remember \
  -H 'Content-Type: application/json' \
  -H "X-Mnemlet-Key: $MNEMLET_API_KEY" \
  -d '{"content":"Christoph prefers dark mode in all editors","namespace":"preferences","memory_type":"preference"}'
{"memory_id":"7f4a...","stored":true,"namespace":"preferences","superseded_ids":[]}

$ curl -X POST http://127.0.0.1:14060/api/v1/ingest \
  -H 'Content-Type: application/json' \
  -H "X-Mnemlet-Key: $MNEMLET_API_KEY" \
  -d '{"content":"token sk-aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa"}'
{"detail":"secret_guard_blocked: patterns=openai_key"}

$ curl http://127.0.0.1:14060/api/v1/audit -H "X-Mnemlet-Key: $MNEMLET_API_KEY"
{"events":[{"action":"ingest","result":"blocked"},{"action":"ingest","result":"success"}]}

$ curl http://127.0.0.1:14060/api/v1/explain/7f4a... -H "X-Mnemlet-Key: $MNEMLET_API_KEY"
{"memory_id":"7f4a...","trust":{"ingested_by":"rest","secret_guard_result":"clean"}}
$ █

Honest Comparison

No checkmark bingo. Here's where Mnémlet shines, and where it doesn't.

	Mnémlet	Mem0	MemPalace	Engram	NeoCortex
Self-hosted	✅	⚠️	✅	✅	❌
Decay / Forgetting	✅	❌	❌	❌	✅
Sleep / Consolidation	✅	❌	❌	❌	❌
Local LLM support	✅	❌	❌	❌	❌
Inspectable vault	✅	❌	❌	⚠️	❌
Vector search	✅	✅	✅	❌	✅
TUI Dashboard	❌	❌	❌	✅	❌
Cloud sync	❌	✅	❌	✅	✅
MCP tools	14	~10	29	19	❌
Pi-friendly RAM	✅ (450 MB)	❌	✅	✅	❌

If your priority is cloud sync, a polished dashboard, or an ecosystem with 50k stars — use Mem0 or MemPalace. If your priority is local-first, brain-inspired forgetting, and running on hardware you own — that's Mnémlet.

Quickstart

GitHub install for now; PyPI remains gated until final release approval.

~/mnemlet — bash

$ pip install git+https://github.com/christoph/mnemlet.git

$ export MNEMLET_API_KEY="mnemlet_..."

$ mnemlet serve

$ curl -X POST http://localhost:4050/api/v1/ingest \
  -H 'Content-Type: application/json' \
  -H "X-Mnemlet-Key: $MNEMLET_API_KEY" \
  -d '{"content":"Hello Mnémlet","namespace":"test"}'

GitHub ↗ Roadmap ↗