One graph memory for every coding agent you use.

One memory, every agent

Use any agent you like. They all remember the same things.

An agent's own memory belongs to that agent. graphersal-memory is one store that every agent on your computer reads and writes: what you settled with Claude Code on Monday, opencode knows on Tuesday, and Cursor next week.

Don't trust us

Prove it on your own work before you switch.

One command, graphersal-memory bench, asks your agent the same questions about your past work with your current memory and with graphersal-memory, scores every answer, and shows quality, cost and tokens side by side. Numbers from your data, not our promises.

What a report looks like

# graphersal-memory bench
harness `claude -p --model sonnet`; 6 question(s) from 9 session(s), 1 repeat(s)

| Category      | Arm   | Mean score  | Confidently wrong | Mean cost | Cost per point |
|---------------|-------|-------------|-------------------|-----------|----------------|
| rules (2)     | files |        4.50 |                 0 |    $0.019 |         $0.004 |
| rules (2)     | graph |        5.00 |                 0 |    $0.022 |         $0.004 |
| decisions (3) | files |        1.00 |                 0 |    $0.017 |         $0.017 |
| decisions (3) | graph |        3.33 |                 0 |    $0.064 |         $0.019 |
| all (6)       | files |        2.00 |                 0 |    $0.018 |         $0.009 |
| all (6)       | graph |        3.50 |                 0 |    $0.046 |         $0.013 |

Spent (list price): $0.91 in all: questions $0.17, warm-up $0.19, answers $0.42, judge $0.14.

## Caveats
- Small sample: 6 question(s) x 1 repeat(s), one project, one model. ...
- The judge is an LLM (the same harness and model as the answers) ...
- File memory holds no session history, so history questions favour the graph arm ...
  ... the full report also lists tokens, turns, time and every answer with the judge's note

A real report, abridged (some rows and columns cut, "..." marks cuts): the first end-to-end run of bench on 2026-11-10, made to check the command (its no-memory baseline arm is cut here), on a copy of one developer's memory with session summaries off. With so few questions it shows the format, not how much memory helps; here the graph scored higher but cost more per point. The report states its own caveats like this every time.

Our own measurement: 21 questions, one real memory

Before the product, we measured the prototype it is built from on Graphersal's own development memory: 21 questions about rules, decisions and past sessions, answered once with the usual memory files and once with graph memory, each answer scored from 0 to 5.

How it works

Search instead of loading everything

Your agent gets a few small tools: search memories, read one, remember, forget, search past sessions. It fetches the few entries that answer the question.

1. Install onceA short wizard sets up every agent it finds and imports the memory files you already have.
2. Work as usualThe agent saves rules, decisions and preferences as you go, per project.
3. Sessions are indexedWhen a session ends, it becomes searchable in the background, whichever agent it ran in.
4. Ask anything later"Why did we drop that?" The agent searches by meaning, across memories and past sessions.

Supported agents

Works with the agents you already use

Install any of them, or all at once: they share one memory. Everything below is pre-release; the status says honestly how far each one has been tried.

Platforms: macOS, Linux and Windows; the whole test suite, end to end included, runs on all three.

Privacy

Your data stays on your machine

Agent sessions can contain anything: code, customer names, a password pasted by mistake. graphersal-memory is designed so that nothing leaves your computer unless you ask.

The only network access of install is the one-time download of the embedding model you choose, from Hugging Face. Pick the built-in model for none at all.