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]"     # the midas-mcp command, for any MCP client
# or, no Python:  npx -y midas-memory-mcp       # TypeScript port
# or, as a library:  pip install "midas-memory[local]"

Install in your agent · See the benchmarks · Team / Enterprise

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planning but blocks memory-justified external or destructive actions unless you explicitly confirmed them — and a superseded memory can't authorize an action at all.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.

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 Continuity Bench scores action-safety, decision-adherence, and repeated-mistake avoidance across a scripted multi-session project (deterministic, no LLM).

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

Midas is a standard MCP server: every client launches the same midas-mcp command with a few env vars — only where you put the config differs. The universal block:

{
  "mcpServers": {
    "midas": {
      "command": "midas-mcp",
      "env": {
        "MIDAS_MCP_EMBEDDER": "local",
        "MIDAS_MCP_DB": "/home/you/.midas/memory.sqlite3",
        "MIDAS_MCP_MAX_RECORDS": "50000",
        "MIDAS_MCP_MIN_IMPORTANCE": "2"
      }
    }
  }
}

Claude Code (CLI, no file editing):

claude mcp add midas -s user \
  -e MIDAS_MCP_EMBEDDER=local -e MIDAS_MCP_DB="$HOME/.midas/memory.sqlite3" \
  -e MIDAS_MCP_MAX_RECORDS=50000 -e MIDAS_MCP_MIN_IMPORTANCE=2 -- midas-mcp

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

Point Claude Code, Claude Desktop, Cursor… at the same MIDAS_MCP_DB file and they share one live memory — 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. Scope it 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.}

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"})

Commercial path

The core stays open source under Apache-2.0: local SQLite memory, MCP tools, SDKs, and the eval harness are free to use, fork, and embed. Paid work is focused on what teams need around that core:

The commercial line is deliberate: Midas does not monetize by closing the memory core. It monetizes operational trust, deployment support, team controls, and benchmark-grade evaluation.

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. Full details in PRIVACY.md · Apache-2.0.