Midas — governed memory for AI agents
A memory your agent can be trusted to act on. Local, source-traceable, and mechanically blocked from acting on stale or unconfirmed beliefs.
Why Midas
Private by default
No LLM at ingest. Your memory stays local unless you choose otherwise.
Safe to act on
A guard blocks stale, superseded, or unconfirmed beliefs before actions.
Auditable
Every memory keeps source, time, provenance, and revision history.
Current Signal
Beyond the hero numbers: second-dataset recall, frontier benchmark, ingest cost, and precision.
See & control your memory.
A local inspector shows what your agent remembers, where it came from, and whether it is safe to use.
- Search every memory with provenance.
- See belief history and revisions.
- Check whether memory can authorize an action.
- Forget with an erasure receipt.
$ midas inspect → http://localhost:7777 · local only # 3,412 memories · belief revisions: 184 · pinned: 12 # governance: 7 actions blocked this week (stale or unconfirmed) ? launch_date "Sept 14" pinned · confirmed 2026-06-12 ? launch_date (prior) "Sept 7" superseded 2026-06-12 ? api_key_rotation_sla "24h" source: ops/policy.md#L42 forget "draft pricing v1" → receipt: sha256:9f2c…b71 ok
Ingest
Deterministic, local, source-aware.
Store
Versioned memory with provenance.
Recall
Relevant context, ready for prompts.
Guard
Blocks unsafe memory-based actions.
Install the SDK
One command. No accounts, no keys, no telemetry. Runs locally against a SQLite file.
Read the install guide →pip install "midas-memory[all] @ git+https://github.com/vornicx/Midas"
from midas import Memory, LocalEmbedder
mem = Memory(embedder=LocalEmbedder())
mem.remember("Launch is on Nov 12.", kind="fact", importance=5)
print(mem.build_context("When is launch?", token_budget=128))import { MidasClient } from "midas-memory";
const mem = new MidasClient({ db: "~/.midas/memory.sqlite3" });
await mem.remember("Launch is on Nov 12.", { kind: "fact", importance: 5 });
console.log(await mem.buildContext("When is launch?", { tokenBudget: 128 }));SDK + MCP
Integrate Midas into your agents and workflows with a simple API or via MCP.
View SDK docs →from midas import Memory, LocalEmbedder, ContentImportance
mem = Memory(embedder=LocalEmbedder(),
importance_scorer=ContentImportance())
# Ingest — LLM-free, deterministic.
mem.remember("Quarterly update: revenue grew 23%.",
kind="fact", importance=5)
# Retrieve — source-traceable, ranked.
hits = mem.recall("What drove revenue growth?",
limit=5, filters={"time_range": "30d"})
# Assemble a budgeted, prompt-ready context block.
context = mem.build_context("When do we launch?",
token_budget=128)
print(context)- Claude DesktopReady
- CursorReady
- VS CodeReady
- Custom AgentsHTTP / SSE
Use cases
Memories that power better agent outcomes.
Support agents
Keep customer context across long conversations.
Research agents
Carry findings and sources across sessions.
GTM agents
Remember account context without leaking data.
No subscriptions. Ever.
Midas is fully open source under Apache-2.0. Everything — engine, guard, inspector, benchmarks, MCP server, CLI — lives in the repo and stays free.
Free forever
No accounts. No paywalls. No usage limits. Just clone, install, and run.
Local by default
Your memory is a SQLite file on your machine. No telemetry, no cloud, no phoning home.
Apache-2.0
Use it in personal, commercial, or on-prem deployments — the license won't bite.
Built to pass a security review.
Shared memory for agent fleets without giving up control. Self-host it inside your perimeter — same Apache-2.0 code, no vendor in the loop.
Scoped memory
Roles, scopes, and team isolation.
Audit trail
Know what memory authorized or blocked.
Provable forgetting
Receipts prove deletion, not hiding.
On-prem / VPC
Deploy inside your perimeter.
Short answers to the hard questions.
Where does my data live?
With you: local file, your server, or your VPC.
How is Midas $0 at ingest?
It does not call an LLM to store memory.
What stops an agent from acting on a stale fact?
The guard refuses stale, superseded, or unconfirmed memory.
Can I really delete a memory?
Yes. Forgetting returns an erasure receipt.
Built in the open.
Midas is developed on GitHub. No sales funnel, no waitlist — just the repo, the docs, and the community.
Prefer email? Reach out at hello@archic.es.
Docs, Benchmarks,
and Comparisons
Everything you need to evaluate and build.