Local-first memory for agents that remember.
Midas is an alpha Python SDK and MCP server for durable, source-traceable agent memory across multi-day coding, research and operational work.
Usable alpha, measured first.
Reader-independent recall@k on reproducible local runs. Ingest stays LLM-free by design.
Benchmarked where memory can be isolated.
Midas leads with deterministic recall@k and structural ingest cost, not demo anecdotes. The alpha ships with the eval harness, reproduce commands and honest caveats in the repo.
A memory layer, not a promise deck.
Midas is the active Archic wedge: one concrete SDK that agents can run today, with local-first defaults and measurable behavior.
Memory that remembers on its own.
Install the alpha as a library, or run it as an MCP server and your agent starts remembering by itself — scoring importance, skipping trivia, with no extraction LLM in the ingest path.
from midas import Memory
# One file. No server, no cloud, no LLM at ingest.
mem = Memory(path="./agent.midas")
# Capture turns deterministically — no LLM extraction.
mem.remember(
text="User prefers concise, technical answers in Spanish.",
source="settings/preferences.md",
)
# Recall with provenance you can audit.
hits = mem.recall("what tone should I use?", k=4)
for h in hits:
print(h.score, h.source, h.span)
# Assemble a context block ready for the next LLM call.
ctx = mem.assemble("plan the next sprint", budget_tokens=1200)# Run the MCP server — any MCP client can call it. midas serve --path ./agent.midas # Tools exposed: # memory.remember memory.recall # memory.assemble memory.forget
Three verbs.
remember() persists. recall() retrieves with provenance. assemble() builds a context block ready for the next LLM call.
Supported: Claude Code · Cursor · Codex · Windsurf · Claude Desktop
Alpha means usable, typed and measured.
The repo includes the SDK, SQLite persistence, local embedders, MCP tools, LangGraph integration, retention primitives, tests and benchmark commands.
Most agent memory systems treat the LLM as the storage primitive. They extract "facts" at write time, re-rank them at read time, and pay token cost on every turn. The result is a memory that is expensive, non-deterministic, and impossible to audit — the model decides what mattered, and you take its word for it.
Midas inverts the stack. Persistence is a single SQLite file you can cp, version, encrypt, or ship inside a container. There is no service to operate, no managed vector DB to budget for, no outbound network at ingest.
Every claim on this site is reproducible. LongMemEval-s recall@k of 0.95 and LoCoMo 0.85 are not screenshots — they're a harness in the repo you can point at your own corpus. Memory systems without published numbers are not engineering, they're marketing.
When you call remember(), no model is invoked. Text is chunked, embedded, and indexed deterministically. The same input always produces the same store.
Alpha, not vapor.
The page points to what exists today: install, run, remember, recall, maintain and benchmark. The roadmap stays narrow because the product is proving memory before expanding the platform.
Benchmarking
Detailed runs across LongMemEval-S and LoCoMo datasets, with reproducible commands.
Architecture
The underlying SQLite + local-embedding implementation, retention tiers and provenance.
MCP Protocol
How to connect Midas to any Model Context Protocol host — Claude, Cursor, Codex, Windsurf.
Research
Open research frontier on calibrated uncertainty, selective forgetting and consolidation.
Midas vs the field.
Head-to-head with mem0, Letta (MemGPT), Zep and LangMem — feature tables, trade-offs and when to pick each one.
See all comparisons →