The SQLite of agent memory. memd is an embedded-first memory engine for AI agents: one process, zero external services. Raw + fact lanes, bitemporal supersedence, provenance and trust tiers, hybrid retrieval and budget-aware packing, in a library that owns a directory. Apache-2.0.

mem = Memory("./my-data")    # a directory, not a service

Status: alpha (0.3.x). See what it does not do yet.

• Embedded, zero keys. No database to provision, no server, no account. Without API keys memd runs on deterministic hash embeddings (or local ONNX embeddings with memd[local-embeddings]) and pattern-based fact extraction, and stats() says which. Bring an OpenAI-compatible key for real embeddings or LLM fact extraction.
• No model call on the write path. A write is acknowledged once it is durably appended to the namespace's log; embedding happens in the background, and the record is searchable by BM25 as soon as the call returns.
• Raw and fact lanes. What was said is kept verbatim; facts are saved explicitly (remember) or extracted when a session closes, and extraction can be re-run because the raw lane is kept.
• Revisable facts. A new fact on an entity key supersedes the old one without deleting it: get(id, history=True) walks the chain and search(as_of=...) asks what was true at a past time.
• Provenance and trust. Every record carries its source tier (user > agent > tool > web > import). Untrusted content is fenced as data in the packed context, explicit saves inherit a session's taint, and bursts of writes or repeated near-identical content are quarantined.
• Hybrid retrieval. A rules-based planner fans out over BM25 (SQLite FTS5, optionally accelerated by tantivy), entity, time and vector lanes, fuses them with reciprocal rank fusion, optionally reranks, and packs the result into a token budget in a stable order.
• Deletion that holds. Hard delete with a physical-purge deadline, forget-by-query with a preview, per-namespace crypto-shred, and a hash-chained audit log.
• Grows without an API change. The same Memory runs on an s3:// root (S3, R2, MinIO), and several server processes on one bucket split the namespaces between them, with keys held by AWS KMS or Vault transit.

memd is not on PyPI yet. Install it from this repository (Python >= 3.11):

pip install "memd @ git+https://github.com/siinghd/memd.git"

or from a checkout, with the extras you need:

git clone https://github.com/siinghd/memd.git && cd memd
pip install -e ".[mcp,s3]"     # extras: mcp, s3, fast, ann, local-embeddings, jev, billing

from memd import Memory

mem = Memory("./my-data")
mem.add("We deploy with `make ship`, never CI", user_id="u1", session_id="s1")
mem.remember("The user prefers dark mode", user_id="u1", entity_keys=["user.theme"])

hits = mem.search("how do we deploy?", user_id="u1", budget_tokens=500)
print(hits.items[0].content)   # We deploy with `make ship`, never CI
print(hits.packed_context)     # provenance-tagged, ready to put in a prompt
mem.close()

In an agent loop it is two calls: messages = mem.pack(messages, user_id="u1") before your LLM call, mem.observe(messages, response, user_id="u1") after. Restart the process and ask again: the memory is in ./my-data. More in examples/.

• Python: from memd import Memory, embedded as above; Memory(api_key=..., base_url=...) is the same API over REST.
• MCP: memd serve --mcp exposes four tools (memory_search, memory_save, memory_forget, memory_status) to any MCP client (setup).
• HTTP: memd serve --http serves the REST API on port 8700, with per-namespace API keys (memd key create --namespace acme).
• TypeScript: sdk-ts/ is @memd/client, a typed REST client for Node >= 18, Bun, Deno and edge runtimes (not on npm yet: build it from sdk-ts/).

Session retrieval on LongMemEval_S through the public Memory.search (dev-fold means ± fold std, from BENCHMARKS.md):

This is retrieval quality on one public dataset. The end-to-end QA numbers are preliminary and the one-time held-out 500-question run has not been done; how these were measured and what they do not show is in BENCHMARKS.md.

• One writer per namespace. A second process opening the same namespace fails fast with NamespaceBusyError; scale by spreading namespaces across processes. Several writers inside one namespace and read replicas are not built.
• Fact extraction is pattern-based by default. An LLM extractor is available with your own key; it has not been evaluated.
• Alpha software. What changed in each release is in the changelog.

• README-engine.md: the full engine guide (running it, S3, key custody, multi-node, hosted mode, retrieval options, ops)
• Docs site: https://siinghd.github.io/memd/ (source in docs/; build it locally with pip install -e ".[docs]" && mkdocs serve)
• BENCHMARKS.md: how quality and latency were measured
• SECURITY.md: the threat model, what it does not cover, and how to report a vulnerability
• CONTRIBUTING.md: setup, tests, and what a change needs
• CHANGELOG.md: every release

Apache-2.0.