Midas

The local memory layer for long-horizon AI agents — remembers across sessions, keeps what's current, and won't act on stale memory.
No LLM at ingest · $0 per message · fully local · every recall traces to its source.

Your AI assistant forgets everything between sessions. Midas is the memory that lives next to it, on your machine. Your coding agent remembers the decisions, conventions, and bugs from three sessions ago — without piping every message through an LLM to "extract" facts. It costs nothing per message, nothing leaves your computer, every memory traces back to the exact turn it came from, and it won't let an agent act on memory that's stale or never confirmed.

uv tool install "midas-memory[mcp,local]"   # install
midas init                                  # create the shared memory + wire up your MCP clients
# or, no Python:    npx -y midas-memory-mcp     # TypeScript port
# or, as a library: pip install "midas-memory[local]"
uv tool upgrade midas-memory   # Update Midas

Install in your agent · See the benchmarks · Complete overview · Roadmap

Why Midas

Most memory tools call an LLM to summarize every session — so you pay in tokens forever, add latency, ship every turn to a provider, and get back rewritten facts you can't audit. Midas makes the opposite bet, and that bet is what makes it cheap, private, and trustworthy:

• $0 and private by construction. No LLM at ingest or query → no API spend, nothing leaves your machine, fast local ops (~tens of ms, no per-turn network round-trip).
• You can trust what it recalls. Recall returns the verbatim source turn, not an LLM rewrite — so there's no extraction step that can silently hallucinate a "fact" you never said.
• It stays current on its own. Typed belief revision supersedes the old value instead of piling up duplicates; selective forgetting keeps it bounded — all with no LLM.
• It's safe to build on. A provenance guard lets memory inform planning but blocks memory-justified external or destructive actions unless you explicitly confirmed them — and a superseded memory can't authorize an action at all.
• One file, many tools. Point Claude Code, Cursor, and your chat app at one SQLite file and they share one live memory.
• Proven, not asserted. Every claim has a reproducible benchmark — including the experiments that failed.

How Midas compares

Every Midas number below is measured and reproducible from this repo; the LLM-at-ingest column reflects the structural properties of that design class (Mem0, Zep, Hindsight) and the figures documented in BENCHMARKS.md.

The last row is deliberate: Midas trades whole-conversation abilities for $0, privacy, and auditability, and publishes the measurements that show exactly where that trade bites.

More than recall: a memory you can govern

Finding a buried fact is table stakes. A long-horizon coding agent needs memory it can act on safely and resume from cleanly — which is where similarity search alone falls short:

These properties are measured, not asserted — the agent-memory bench suite scores action-safety, decision-adherence, repeated-mistake avoidance, resume fidelity, conflict detection/precision (live contradictions between agents found without over-flagging), and adversarial memory-safety across scripted multi-session projects. The safety eval blocks 10 / 10 adversarial attacks (ASR 0.00) — including a planted confirmation next to a prohibition, a confirmation for a different action, a provenance-laundering supersession, and a cross-namespace approval — with no over-blocking (benign-pass 1.00). Deterministic, $0, no LLM. Reproduce every number with one command:

uv run python -m eval.benches      # the whole governance suite — or `midas bench` from a checkout

How it does on the benchmarks

Deterministic, reader-independent retrieval (recall@k — fraction of the gold supporting turns pulled into context) on the full public sets, vs a recency-window baseline:

And the cross-system metric, judged answer-rate (same gpt-4o judge the leaderboards use):

All of it at 0 LLM calls, $0, and 0 data egress at ingest. Full numbers, per-category breakdowns, reproduce commands, and the head-to-head vs Mem0/Zep/Mastra are in BENCHMARKS.md.

Eval-first means we publish the misses too. Hybrid retrieval, reranking, thread-diversification, dual-granularity indexing, and naive distillation were all measured to not help (or to hurt) and are documented as such. That honesty is the point — see BENCHMARKS.md and docs/frontier-2026.md.

Connect it to your coding agent

One command wires up everything:

midas init        # creates the shared memory + configures every MCP client it finds
midas status      # check what's wired   ·   run `midas init --dry-run` to preview first

Both take --json to emit a machine-readable client wiring receipt — which memory each client got wired to, under which scope/policy, and which clients were skipped (config paths only, never memory contents). Paste it into a bug report, or let another agent verify the setup without scraping prose.

midas init creates one shared memory (~/.midas/memory.sqlite3) and points the MCP clients it detects — Claude Code, Codex, Grok Build, Cursor, Claude Desktop, Windsurf, VS Code, Gemini CLI, Cline, Zed — at it. So all your agents read and write the same memory, autonomously, with no per-client paths to keep in sync.

Prefer a single endpoint over per-client launches? Run one server and give your clients an MCP URL:

midas serve --http        # → http://127.0.0.1:7077/mcp   (one server, one memory, every client shares it)
midas serve --http --token <secret>   # require `Authorization: Bearer <secret>` on every request

Keep Midas current with midas update. See your memory anytime with midas inspect.

Already carrying agent memory in files? midas import --from claude-md CLAUDE.md (or --from cursorrules, --from jsonl, --from mem0, --from zep) turns those rules and exports into first-class, recallable, governable memories — tagged with where they came from, idempotent on re-run.

Want memory even when the agent never calls capture? midas init --claude-hook installs a Claude Code SessionEnd hook that offers each session's user turns to memory — Midas's no-LLM policy still decides what is actually kept.

{ "mcpServers": { "midas": { "command": "midas-mcp", "env": { "MIDAS_MCP_EMBEDDER": "local" } } } }

Once connected, Midas injects a short policy into the agent (recall first, then capture durable facts/decisions/preferences/constraints/corrections). The agent captures freely; Midas decides what's kept — it scores importance (no LLM), drops trivia, skips duplicates, revises stale beliefs, and forgets the low-value tail to stay bounded. Before any memory-justified external or destructive action, the agent calls check_memory_use and is blocked unless you confirmed it (and that confirmation is still current).

One memory, many clients

By default every client shares one live memory (~/.midas/memory.sqlite3) — each detects the others' writes (SQLite data_version) and refreshes, so a fact captured in your IDE is recallable from your chat app seconds later, no restarts.

Want per-project separation instead? midas init --project-scoped (or MIDAS_MCP_NAMESPACE=auto) gives each project its own partition in the same store — the scope is derived from the git repo / cwd the server runs in. Or scope it manually per project/agent/user with MIDAS_MCP_NAMESPACE.

_{Real run, reconstructed chrome — the capture/recall lines are verbatim output of two separate processes sharing one file.}

Troubleshooting

Something not wired right? midas doctor is the one-command diagnosis — it checks midas-mcp is on PATH, that your store opens, whether the local embedder is available, and which clients are actually wired, with a fix hint per failed check. It reads config paths and versions only — no memory contents, so its output is safe to paste into a bug report.

midas doctor          # ✓/⚠ per check, with a hint for each failure
midas status          # what's wired + the store's record count

Still stuck? Open a bug report (it pre-fills the midas doctor block) or ask in Discussions.

Use it from Python (the SDK)

from midas import Memory, LocalEmbedder

mem = Memory(embedder=LocalEmbedder())   # fully local. (Or Memory() for a zero-setup offline embedder.)

mem.remember("Decision: the primary database is PostgreSQL.", kind="constraint", importance=5)
mem.remember("The launch date moved to September 14.", kind="fact", importance=5)
mem.capture("lol ok cool")               # filler — auto-scored below the floor, skipped (no LLM)

mem.assemble("when do we launch?", token_budget=128)          # prompt-ready, dated, source-traceable
for hit in mem.recall("which database did we pick?", limit=3):
    print(f"{hit.score:.2f}  {hit.record.content}")           # each hit traces to its source

from midas import Memory, LocalEmbedder
from midas.nli import LocalNLI
from midas.sqlite_store import SQLiteStore
from midas.state import memory_state, memory_diff   # the control-plane views

# Durable, shareable, no native extension. Safe across threads & processes (live data_version refresh).
mem = Memory(store=SQLiteStore("memory.db"), embedder=LocalEmbedder(),
             supersede=True, nli=LocalNLI())   # a turn that CONTRADICTS an old belief supersedes it

# Control-plane: the current state of a project, and what changed since a point in time (no LLM):
memory_state(mem, scope={"project": "apollo"})          # live, non-superseded decisions/constraints/facts
memory_diff(mem, since=last_session_epoch)              # {added: [...], revised: [(old, new), ...]}

mem.forget_decayed(max_records=50_000)         # evict lowest value (importance × recency); protects facts
mem.recall("when is the launch?", as_of=1_700_000_000)   # bitemporal: "what did we believe on date X"

# Right-to-be-forgotten — preview, then erase, with an audit trail:
mem.forget_matching("the user's home address", dry_run=True)
mem.forget_matching("the user's home address")

# Back LangGraph's long-term memory with Midas:
from midas.integrations.langgraph_store import MidasStore
store = MidasStore(); store.put(("user", "123"), "pref", {"text": "prefers dark mode"})

See & control your memory — midas inspect

Most memory is a black box of LLM-rewritten facts. Midas is glass-box: run a local inspector over your store and see exactly what your agent remembers, why, and from what source — then correct, pin, or forget it.

midas inspect --db ~/.midas/memory.sqlite3      # opens http://localhost:7777 — local only, zero egress
# before install:  python -m midas.inspector --db <your.sqlite3> --embedder hashing

• Overview — counts, attributability, a 30-day activity chart, and kind/provenance/recency breakdowns, each kind and provenance color-coded consistently across every view (a fixed categorical palette, validated for colorblind-safe contrast in both themes — never color-only, every value keeps its label).
• Browse + search every memory (verbatim, with provenance + source), filterable by kind, provenance, and sort order.
• Belief history + time-travel — what you believed, what it superseded, and when.
• Project state (decisions / bugs / forbidden) and what changed since a date.
• Governance — would memory authorize an action, and why (the audit trail); forget with a receipt.
• Conflicts and Open loops — the same control-plane views from memory_conflicts/open_loops, with one-click resolve/close from the UI.
• Audit log — the hash chain's verification status and its most recent entries.
• Light + dark themes (a real second theme, not an inverted dark one), keyboard shortcuts (⌘K to jump anywhere or search, / to focus search), and a responsive layout down to phone width.

And every mutation (write / revise / forget) appends to a tamper-evident, hash-chained audit log inside the store — hashes only, never content. midas audit shows it; midas audit --json verifies the whole chain and reports the first broken entry if anyone rewrote history.

No LLM, no account, runs on your file. The thing a black-box memory can't show.

Free & open source

Midas is fully free and open source under MIT — the memory engine, the guard, the MCP server, the CLI, the inspector, the TypeScript port, and the entire bench suite. No paid tiers, no feature gates, no telemetry, no account. Use it, fork it, embed it in commercial products — the license permits all of it.

If Midas is useful to you, the best ways to give back are a ⭐, a reproduced benchmark number, a bug report, or a measured contribution — see the roadmap for where help matters most.

Honest status

Midas is early but built narrow and measured-first. Where it stands, plainly:

• Retrieval is its strength and is essentially maxed for a no-LLM design — confirmed by our own A/Bs and by the frontier papers (the retriever is not the bottleneck). The benchmark numbers above are the result.
• The frontier's extra lever is structure-preserving extraction — and it needs a capable model Midas deliberately won't run at ingest. We built the judged harness and measured it on BEAM's summarization category: a small local extractor doesn't help (raw 0.28 vs replace 0.07 rubric coverage), and the lift is gated on a strong model — so it belongs to the agent's model, not Midas's. The optional distillation dial ships off by default; we don't claim it as a win. (Details: docs/frontier-2026.md §2b.)
• Where it's heading: from recall to a governed memory control-plane — memory_state / memory_diff, the provenance guard that won't act on stale or unconfirmed memory, and the Agent Continuity Bench that measures those properties. Local, auditable, and honest about what's proven.

The eval harness

eval/ (dev-only) runs Midas and competitors through synthetic / LoCoMo / LongMemEval / multiday / conflicts-v1 / BEAM with deterministic recall@k + precision@k, cost/latency instrumentation, a dumb-reader ablation (proves the numbers aren't reader-inflated), and an optional local-or-hosted LLM judge. The anti-cheating checklist (no query rewriting, no LLM at ingest, no gold leakage, seeded sampling), conflict handling, failure traces, and the verbatim MCP policy are in docs/methodology.md.

python -m eval.runner --dataset longmemeval --variant s --local --midas-no-rerank --max-questions 40
python -m eval.runner --dataset beam --beam-tier 100K --local --dumb-reader   # frontier benchmark
python -m eval.continuity                                                      # Agent Continuity Bench

Privacy & license

Local-first: every memory lives in a SQLite file on your machine, recall returns the exact stored text, and capture/recall/forget make no network calls. No account, API key, or telemetry. The only outbound traffic is a one-time embedding-model download (for the local backend) and the package install. Optional encryption at rest: set MIDAS_MCP_KEY with the [encrypted] extra and the store is a SQLCipher database — unreadable without the key (and Midas fails closed rather than silently writing plaintext). Full details in PRIVACY.md · MIT.