Memmy

A local memory hub that several coding agents read from and write to. One service — http://127.0.0.1:18960, its data in ~/.memmy/memory-service/memory.sqlite — holds the memory, and hooks, plugins, the memmy-memory CLI and the desktop app all go through it, so Claude Code, Codex, Cursor, DeepSeek Harness, OpenClaw, Hermes, OpenCode, WorkBuddy and Pi can share one store. Each host is connected one of three ways: a hook and workspace bridge written into its own directory, a plugin in its own format, or a read adapter that scans the session files the host already keeps. Memory is layered — L1 traces, L2 policies, an L3 world model and skills — and promoted between layers by measured gain, with reward propagated back through an episode after a 30-second feedback window. Built by the MemTensor organisation in 76 days: 1,279 commits, 20 releases, 2,439 files, 2,000 stars.

Who built itAn organisation account rather than an individual one: MemTensor ships Memmy as a product, with staff addresses on memtensor.cn and memtensor.com, a site at memmy.bot, a Discord server and an X account. Its 1,279 commits come from 25 contributors, and one of the contributing accounts is named Memtensor-AI. The largest single committer is Wang-Daoji with 223 commits, followed by hijzy at 163, syzsunshine219 at 130 and ZongYue99 at 118. Of the 1,279 commits, 921 are linked to an account and 375 are not. The organisation is not the main writer: 153 commits carry a co-author trailer, 54 of those saying Cursor and 64 naming a Claude model, with another 35 naming accounts that also appear in the contributor list.

1. 01

   One local store, ten hosts, twenty releases in eleven weeks

   The repository was created on 2026-07-16 and its first commit is dated 2026-07-17: memmy-agent: Let every AI remember the same you. Version v1.0.1 shipped the same day, and by 2026-09-30 the line had reached v1.2.0 — twenty releases in eleven weeks, the longest quiet stretch being the nine days between v1.1.1 on 2026-08-26 and v1.1.2 on 2026-09-04. The commit curve is 240 in what was left of July, 450 in August and 589 in September, 1,279 in all across 76 days. Around the code sit 2,000 stars, 194 forks, 5 watchers, 37 open issues and 25 contributors, and the README carries a Product Hunt daily top-post badge. The claim the project makes for itself is in the repository description — in its words, “all AI remember the same you” — and the mechanism is a single service rather than one store per tool: the endpoint is http://127.0.0.1:18960, the data is ~/.memmy/memory-service/memory.sqlite, and the documentation states that hooks, plugins, the memmy-memory CLI and the desktop app all read and write that one service, which is what lets different agents share the same memory.

2. 02

   Three ways into a host, because no two hosts agree on anything

   A memory hub is only useful if it can get in and out of other tools, and the repository shows three distinct routes. The first is an installed target per agent: Memory/src/agent-source/integration/ holds one directory per host — claude-code at 13.6 KB, codex at 12.4 KB plus a 9.6 KB hook-trust.ts, cursor at 9.0 KB, deepseek-harness at 7.9 KB, opencode at 7.5 KB, and much larger ones for hermes at 78.8 KB and openclaw at 64.3 KB, with pi and qwenwork under 600 bytes each — alongside templates that emit what gets written: a 43.8 KB resume hook, a 39.0 KB OpenCode plugin and a 25.2 KB DeepSeek Harness plugin. The second route is a plugin in the host’s own format: Memory/adapters/openclaw/openclaw.plugin.json, a Cordis patch file for DeepSeek Harness, and a Python memmy_provider package for Hermes. The third route only reads: one adapter per host that scans the session files a tool already keeps. How little uniformity there is between them shows in those adapters — Cursor’s workspace state is a SQLite file, Codex keeps rollouts, OpenClaw and OpenCode have databases of their own, and DeepSeek Harness writes compressed session.jsonl.zstd files.

3. 03

   What controlled and local-first mean in the code

   The service can be installed without touching any agent: memmy-memory install --service-only downloads the pinned runtime, registers a user-level service, starts it and checks /api/v1/health, and the README says it does all of that without installing a skill or adapter anywhere. The Linux installer initializes memory and leaves Codex, Claude Code, Cursor and the rest alone until the user runs memmy-memory init, or memmy-memory init --agent <agent>. Two settings turn memory reading and writing off independently — enableMemorySearch and enableMemoryAdd — and the documentation notes that neither forces the other. Configuration is a single file at ~/.memmy/config.yaml, and recall, evolution and model settings can be reloaded with memmy-memory reload-config; changing storage instead returns requiresRestart: true rather than pretending to have applied. The surface for checking what the machine actually knows is small and inspectable: memmy-memory search "..." --verbose prints the raw candidates gathered by each channel, the hits left after fusion and thresholds, the memories really rendered into the agent’s context, and whether the model filter ran, was skipped or fell back.

4. 04

   Four layers, six channels, and a two-hour definition of a conversation

   The memory is layered, and every promotion between layers has a number attached. turn.complete writes a raw turn and one L1 trace per captured step, keeping up to 4,000 characters of text and 2,000 per tool output; consecutive follow-ups merge into one episode with a maximum gap of two hours; when an episode closes the service scores a reflection, waits a 30-second feedback window, computes a task reward and propagates it back through the episode’s traces with a 0.9 gamma and a 30-day decay half-life. L1 traces above a value of 0.005 and a similarity of 0.65 become L2 policies when their gain clears 0.02, and are archived below -0.05; clusters of L2 at 0.3 similarity become an L3 world model that needs a confidence of 0.2 to be recalled; skills crystallize from policies at an eta of 0.1. Recall uses four families — vectors, SQLite FTS5, a short-pattern channel built for two-character Chinese fragments and short ASCII terms, and structural fragments for error signatures and paths — over three tiers sized 3, 5 and 2, which is also the default limit of ten results, then MMR at 0.7 relevance against 0.3 redundancy, then an optional model filter that keeps at most eight and falls back to six when it fails. Vector search builds its window from the most recent 2,000 rows, and the documentation says that number is a fixed constant rather than an option.

5. 05

   Release branches with the boundaries written into the pull request

   The release process is unusually explicit for a project this young. Work lands on a version branch, and the pull request that cuts a release states what it will not do: a v1.1.9 candidate in September carries a section headed “Release boundaries” — draft pull request only, no version tag pushed, no GitHub release published, no installer uploaded. Two candidates were abandoned and superseded within the day, one of them for targeting the version branch instead of main. The merge pull request for v1.2.0 carries three review handoff markers, each an HTML comment holding a hash and a timestamp, with the note that only reviews submitted strictly after that timestamp qualify. The task tracker is not GitHub: work arrives as identifiers such as MEMMY-491, MEMMY-487, MEMMY-607 and MEMMY-623, which the pull requests describe as tasks in Yunxiao and which a workflow named for Yunxiao mirrors onto GitHub, so requirements are filed in one system and code lands in another. Release notes live in the repository and shorten as the pace rises — 4,731 characters for v1.1.2 in early September gave way to 455 for v1.1.8 and 695 for v1.2.0 — and one candidate PR includes a fix to two stale regression assertions that still pinned the backend to 1.1.7 while the root version was already 1.1.8.

6. 06

   A numbered bug list, synthetic samples, and what the fixes cost

   Most contributors come from outside the organisation and work from a numbered list of known problems; the pull requests cite entries such as item 2, item 5, item 39 and item 60, and most of those fixes are still open rather than merged. An analytics call was sending the search query, the memory body and the memory ID to a cloud endpoint as the “action” name — the skill this project installs into agents teaches those exact forms — and the fix reports a known subcommand or nothing. Closing a session woke the background worker only when the response happened to carry a closed episode, so newly frozen L3 jobs sat; the pull request measures the wait at 2 minutes 41 seconds and replaces it with one unconditional call. read_file counted a trailing newline as an extra line while apply_patch did not, so a model patching a one-line file wrote a diff that could not apply. Turn capture excluded every turn the relation model labelled as ending a topic, discarding a user’s feedback when it shared a sentence with a goodbye; only a bare end command is excluded now. The samples in these pull requests are synthetic and the authors say so, and where testing stopped short they say that too: the change adding subscription plans as providers has not been called with a real key. An offer of a hosted namespace behind the local hub, filed as issue 547, has no reply.