Memory for AI agents that stays on your servers — and tells you when it's lying.

Live proof

Same prompt, safer outcome when memory is auditable.

Mnemos helps your system avoid confidently repeating stale or contradictory memory by forcing evidence and contradiction handling into the loop.

"Deployment succeeded in production."

No contradiction surfaced.
No evidence shown.
Wrong state repeated.

"There are contradictory claims about deployment status.
Latest resolution: rolled back at 14:02."

evidence: ev_1
contradiction: [cl_a, cl_b]
confidence: 0.84

Fast activation

Start with one endpoint, expand only when needed.

Mnemos exposes one HTTP API. Any language with an HTTP client can add memory in five lines — copy/paste, then own it.

import httpx, uuid
m = "http://localhost:7777"
run = str(uuid.uuid4())

# Remember something
httpx.post(f"{m}/v1/events", json={"events": [{
    "id": str(uuid.uuid4()),
    "run_id": run,
    "source_input_id": "chat-session-1",
    "content": "user prefers vegetarian options",
    "timestamp": "2026-05-03T16:00:00Z",
    "metadata": {"role": "preference"},
}]})

# Recall it later (months later, same call)
events = httpx.get(f"{m}/v1/events", params={"run_id": run}).json()

Proof of value

When facts conflict, Mnemos surfaces it before your AI response hardens the error.

Beyond the five-line memory write, Mnemos extracts structured claims, links evidence, and detects contradictions — the parts that make replay-by-run-id more than a glorified log. No API keys required for any of the below; add an LLM provider only when you want grounded answers.

$ go install github.com/felixgeelhaar/mnemos/cmd/mnemos@latest

$ mnemos process --text "The deployment succeeded in production.
The deployment did not succeed in production. Response times averaged 45ms."
ingested 1 event · extracted 3 claims · 1 contradiction detected

$ mnemos query "What happened with the deployment?"
{
  "claims": [
    { "id": "cl_a", "text": "The deployment succeeded in production",
      "type": "fact", "confidence": 0.91,
      "evidence": [{ "event_id": "ev_1", "span": [0, 47] }] },
    { "id": "cl_b", "text": "The deployment did not succeed in production",
      "type": "fact", "confidence": 0.91,
      "evidence": [{ "event_id": "ev_1", "span": [49, 95] }] }
  ],
  "contradictions": [
    { "between": ["cl_a", "cl_b"], "kind": "polarity_conflict" }
  ]
}

$ mnemos resolve cl_a --over cl_b --reason "rolled back at 14:02"
resolved · cl_a now valid_to=2026-05-02T14:02:00Z

$ mnemos query --at 2026-05-01 "What happened with the deployment?"
# returns the pre-rollback world — same store, point-in-time view

Where it fits

Most memory tools optimize for convenience. Mnemos optimizes for trust.

Mnemos is one of several ways to add memory to an AI app. The right choice depends on whether you need a hosted SDK, raw similarity search, a notes app, or a substrate you self-host. Mnemos is the last one — and the only one that ships with contradiction detection and replay-by-run-id built in.

Architecture

Ingest · Extract · Relate · Query.

Events are immutable. Claims are derived. Relationships connect claims. Embeddings rerank. Trust falls out of confidence × corroboration × freshness. No magic.

Run it

From install to trust signal in under five minutes.

brew install felixgeelhaar/tap/mnemos
# or: go install github.com/felixgeelhaar/mnemos/cmd/mnemos@latest

mnemos process --text "The deployment succeeded in production.
The deployment did not succeed in production."

mnemos query "What happened with the deployment?"

SQLite ships with the binary. No Docker, no hosted service, no API key. Add a provider when you want grounded answers:

export MNEMOS_LLM_PROVIDER=anthropic
export MNEMOS_LLM_API_KEY=sk-ant-…
mnemos query --llm "What happened with the deployment?"

Inspect the audit chain

mnemos history --kind=claim cl_a   # every state change with reason
mnemos verify cl_a                  # re-check against source events
mnemos export --kind=lesson         # YAML-frontmatter markdown for review

Wrap LangGraph, CrewAI, or any MCP-compatible agent

Mnemos doubles as the audit substrate beneath any AI agent. Each node-or-step posts one event keyed to a single run_id; the full reasoning chain is one HTTP call away weeks later.

cd examples/refund_triage_langgraph
pip install -r requirements.txt
python agent.py --customer-id CUST-42 --amount 245.00

curl -s "http://localhost:7777/v1/events?run_id=<run-id>" | jq

The example wires raw HTTP, no SDK — four lines per node get you a defensible audit trail. Source: examples/refund_triage_langgraph/.

FAQ

Questions teams ask before they replace hosted memory