sodamem 0.1.1

A self-evolving, agentic memory layer for AI agents.

Most memory systems store what you said and stop there — correct today, silently wrong the moment your life changes. SodaMem evolves alongside your agent: facts get superseded instead of overwritten, entity profiles rebuild on demand instead of drifting silently stale, and every answer still traces back to the exact turn it came from. Recall costs zero LLM calls, so the same question gets the same answer every time.

English · 简体中文 · 日本語 · 한국어 · Français · Español · Deutsch · Português

Agent integrations · Benchmark · Quick start · Why another memory layer · Install · Use it from anywhere · Coding tools · Self-hosting · Contributing

Accuracy against estimated API cost per question. The quadrant that matters is up and to the left.

Agent integrations

Full index, including MCP tool schemas and adapter details: integrations/README.md.

Benchmark

92.8% (464/500) on LongMemEval.

Every answer and every retrieved memory is published in benchmarking/artifacts/ — 500 answers verbatim, 8,427 evidence rows. Re-grade them with any judge, or hand the retrieved context to your own reader and see what the number does — swap the reader and the score moves, which is why the artifacts are the point, not just the 92.8%. Neither needs access to anything of ours.

86.88% (1338/1540) on LoCoMo. End-to-end QA accuracy, LLM-as-judge.

No per-question artifacts are published for LoCoMo — no answers, no retrieved context, no run directory. What is published is the LoCoMo section of benchmarking/README.md: the per-category breakdown, the per-conversation spread, provenance and repro steps.

Quick start

This is the Python path. Wiring into an agent framework or MCP client? See Agent integrations. Calling it from TypeScript/Node? See Use it from anywhere. Running it as a shared service? See Self-hosting.

Example

pip install "sodamem[chroma,llm]"

from sodamem import SodaMem
from sodamem.llm import create_provider_from_env      # SODAMEM_LLM_API_KEY etc.
from sodamem.memory.ingest.extractor import FactEventExtractorV2

mem = SodaMem.open("./data", extractor=FactEventExtractorV2(create_provider_from_env()))

mem.ingest(
    [{"role": "user", "content": "Actually I moved from Kauai to Oahu."}],
    user_id="u1", session_id="s1", session_time="2023-05-25",
)

block = mem.build_context("where am I staying?", user_id="u1", token_budget=1000)
print(block.text)        # prompt-ready — zero LLM calls
print(block.citations)   # the exact evidence behind every line of it

SodaMem.open() creates ./data if it isn't there. Only .ingest() needs the extractor — drop that argument for a read-only store and search / build_context work exactly the same.

Nothing about you leaves the machine. No telemetry, no analytics, no callback — the only outbound request the default install ever makes is a one-time download of the 90 MB MiniLM embedding model into ~/.cache/chroma/, and after that it talks to nothing but your disk. Pre-seed that cache and it runs air-gapped.

Why another memory layer

Most memory systems store what you said. The questions that break them are when it stopped being true and where it came from — and those need a data model, not a bigger vector index.

Two of these are worth a closer look — the rest is what the table already says.

Every memory carries its receipt

A retrieved memory is not a floating string. It points at the turn that produced it:

evidence_id  = ev_fact:fact_6ada707b…
support      = "Can you recommend a good beach on Oahu that's not too crowded?"
predicate    = user wants a not-too-crowded beach on Oahu
entities     = location=Oahu | occasion=birthday
source       = session_40 / turn_10          ← the exact turn, not "some chat"
date         = 2023-05-25

FactEvent → SourceSpan → RawTurn is a foreign-key chain, not a similarity score, so "why do you think that about me?" has an answer.

Four time axes, not one timestamp

One timestamp cannot separate "I moved to Chicago last year" from "I will move next year", or express a fact that stopped being true.

Install

Base install pulls pydantic, numpy, rank-bm25, python-dateutil — and a CI gate fails the build if that list grows by accident.

Use it from anywhere

HTTP — add / search / context / answer, plus batch write, supersede, events, metrics, token usage:

curl -H "Authorization: Bearer $KEY" -H "Content-Type: application/json" \
  localhost:8000/v1/context \
  -d '{"user_id":"u1","query":"what do they prefer?","token_budget":1000}'

/v1/context and /v1/search both take a JSON body; /v1/context also answers a plain GET with query params, since it is a pure read. The one Python-only exception is build_context(organizer=...), which runs an LLM-backed organizer over the retrieved set for questions like "list everything you know about me" — /v1/context never accepts one, so the zero-LLM guarantee over HTTP can't be flipped by a request parameter.

SDKs — TypeScript over HTTP (sdk-ts/, zero runtime dependencies, ESM + CJS):

npm i sodamem

import { SodaMemClient } from "sodamem";

const mem = new SodaMemClient({ baseUrl: "http://localhost:8000", apiKey: process.env.SODAMEM_API_KEY! });
const block = await mem.context({ user_id: "u1", query: "what do they prefer?", token_budget: 1000 });

Python talks to the library directly — import sodamem and you are already past the network.

Agent frameworks — LangGraph, CrewAI, OpenAI Agents SDK, Vercel AI SDK. Scope is bound when you construct the tools and never appears in the schema the model sees: a user_id the model can choose is a user_id it can hallucinate.

MCP — 8 tools, including entity_timeline (one entity's history in order, each item still pointing at its source) and explore_memory (walk the graph outward). Six are reads and always available; the two that mutate (add_memories, delete_memory) appear only under SODAMEM_MCP_ALLOW_WRITE=true, which sodamem install writes for you into the client config it generates.

Web console — browse and inspect memories per tenant, shipped in the image.

Coding tools

Step 1. Start the daemon — the one process that owns the stores:

sodamem daemon ensure

Step 2. Wire a client to it:

sodamem install claude-code

Every client gets the MCP tool surface. Four also get hooks, so memory is recalled and retained without the model having to decide to call a tool — which in a coding session it mostly doesn't, because it's busy reading files.

What hooks can do is not uniform, because the hook systems aren't. This is what each client actually supports, and sodamem clients prints the same thing:

Cursor's beforeSubmitPrompt can read a prompt but cannot inject anything (its docs list exactly three events that can, and that isn't one), and neither Cursor nor Codex hands a hook a transcript path — so there is nothing for a retain hook to read. Those two get a project brief at session start and write through the add_memories tool. We don't install a hook that can only ever do nothing.

Three things worth knowing before you run it:

One daemon, many editors. Per-user stores are SQLite without WAL, so exactly one process may open them (ADR 0001 §2). install therefore points every client at a running service by default rather than letting each spawn its own — and if you deliberately choose a local store (--local-store), a second client now refuses to start instead of quietly corrupting the first one's data.

Memories are scoped to the repo. install derives a project_id from the git root (a git worktree resolves to its parent repo, so one branch per task is not one memory bank per task). Narrowing, not partitioning: anything you told SodaMem outside a project still surfaces inside every project, and dropping the key answers "how did I fix this in the other repo?".

Retain needs extraction credentials. Recall is zero-LLM and works without them; storing facts does not. sodamem daemon ensure says so up front rather than accepting every write and failing the job afterwards.

sodamem install claude-code --dry-run      # print what would change
sodamem install cursor vscode zed          # several at once
sodamem daemon status                      # what is actually answering

Existing config is merged, not replaced — other MCP servers, other settings and hand-written TOML comments survive — and the first write of any file leaves a .sodamem-backup beside it.

Self-hosting

One command:

cp .env.example .env      # then set SODAMEM_API_KEY
docker compose up -d

Auth is on by default. docker-compose.yml never sets SODAMEM_AUTH_DISABLED — the server refuses to start if SODAMEM_API_KEY is unset (see server/settings.py), so there is no accidentally-open deployment. Set the key in .env before the first docker compose up.

Run exactly one worker. --workers 1 is a correctness constraint, not a throughput setting: per-user stores are SQLite databases opened without WAL, and two processes writing the same user's store corrupt it. The shipped CMD states it explicitly, and the server takes an exclusive lock on its data root at startup — a second process pointed at the same directory refuses to start with data_root_locked rather than quietly corrupting data. Horizontal scaling needs an external job store first (docs/adr/0001-control-plane-db.md).

Full operations reference — calling the API, admin endpoints, metrics, maintenance, backups, upgrades — lives in docs/self-hosting.md.

Contributing

Bugs, features and PRs are welcome. Read CONTRIBUTING.md first — it describes four rules CI enforces that the code itself does not announce.

Acknowledgements

Early contributions from @sunjiajunsunjiajun and @Lum1104 helped shape the work this project grew out of. Thank you.

License

Apache-2.0. Copyright 2026 FENGRONG WAN.