Polis Memory System
Polis gives your AI agents a long memory of your work. It keeps every prompt and decision word for word, in a record that can’t be quietly rewritten. It organizes that record into a catalog that grows with your projects. And when an agent asks, it answers with the exact record each memory came from.
Prompts, plans, approvals and decisions arrive faster than anyone can file them. Polis records each one word for word as it happens, with no model involved.
When the machine goes quiet, the gardener files each record into a catalog of classes. Your projects are the roots; classes split, merge and go quiet as your work changes.
Follow any branch down and you reach the record itself: numbered, hash-chained to the one before it, and quoted exactly as it was written.
“Use SQLite so prompts and decisions stay on the machine, with a history we can verify.”
polis / storage / decisions
Polis records prompts and decisions as they happen, organizes them into a class catalog, and retrieves evidence with references to ledger sequence numbers.
The structure
Polis separates what happened from how it’s organized, and both from what it has inferred. Each layer points back to the one beneath it. None of them can rewrite it.
The lake
Every prompt you send, every plan you approve, every decision and note — stored word for word in an append-only ledger. Each entry is hashed together with the one before it, so changing a single word in the past breaks every link after it, and polis verify will say where. Writing to the lake takes no model call at all.
The catalog
A tree of classes over the lake, seeded by your repositories. Its leaves are pointers into the lake, never copies — so reorganizing the tree can’t touch the evidence. Without a model configured, new records wait in an inbox, and everything stays searchable.
Observations
Notes the gardener writes about patterns across your work, each citing the records it came from. They’re re-checked on later passes and retire on their own when a record they cite is forgotten. Cold branches can be folded into digests the same way: shorter, but one hop from the source.
The answer pack
When an agent asks, Polis assembles one bounded bundle: the matching part of the catalog, your notes, full-text and literal matches, and optional semantic search — ranked together and capped in size. The calling agent writes the answer; Polis supplies the evidence, each piece cited back to its ledger number.
The gardener
When the machine has been quiet for a minute and a half and the lake has grown, the gardener takes a turn. A model proposes moves; fixed rules decide which ones happen. It has exactly seven moves.
Attach a record to a class — only within the project it came from.
Start a new class. Refused if a sibling’s title is 80% the same.
Lift a sub-topic that has grown up a level. No cycles; four levels at most.
Divide a class holding two subjects. Each half needs three records.
Join duplicates. Never across projects.
Fold a cold branch into a digest that cites its records. Never one you’ve annotated. Reversible.
Mark a decision replaced by a newer one. The old one stays in the lake.
Before and after every reorganization, about 200 frozen retrieval probes test what is still findable. If something that was reachable stops being reachable, the run reverts itself and the subjects involved sit out the next three runs.
Each move is journaled with a copy of what it replaced, so any run can be reversed exactly. The last fifty are kept.
There is nothing to approve. You can tell it to remember, annotate what matters, forget what shouldn’t be kept, and revert any run.
The philosophy
Summaries are allowed to blur. The record isn’t. Anything Polis condenses keeps a citation one step from the original.
“Blurriness lives at the summary level — sourcing stays perfect one hop away.”
Every memory carries where it came from — which project, which record. That is written down, never guessed after the fact.
“Provenance is ground truth — never infer it.”
A prompt is stored and returned exactly as written. Nothing inside it can instruct the gardener or change how memory is organized.
“Captured content is data, never instructions.”
A model makes the catalog smarter. Without one, Polis still captures, files to an inbox, and retrieves.
“Capture, retrieval and filing never depend on a model.”
A SQLite store on your machine. No account, no telemetry, no update check. Model calls go only where you point them.
“The default install talks to nothing but itself.”
The catalog is how memory is arranged, not the only way in. Search reaches every record whether or not it has been filed.
“The pack ranks; the tree organizes.”
Quoted lines are from the Polis source and documentation.
From the project’s acceptance run on 10,000 prompts on an Apple M3 Max, on the current main branch. These are speed and recall checks of Polis itself, not a comparison with other memory systems.
Where it works
Polis speaks MCP. Install it into Claude Code, Codex, Cursor, Windsurf or Claude Desktop, and every one of them reads from — and writes to — the same memory. Or run polis serve for a local HTTP daemon, or embed its seven Rust crates in your own tools.
It’s also built into Redline Meta-Harness, where it grew up.
Read
memory_searchmemory_contextmemory_grepmemory_treememory_nodememory_timelinememory_statsmemory_verifyWrite
memory_remembermemory_ingestmemory_annotatememory_supersedememory_forgetPolis is a pre-release. Prebuilt binaries for macOS, Linux and Windows are on GitHub with checksums; Windows installs with PowerShell.
Install guidecurl --proto '=https' --tlsv1.2 -LsSf \
https://github.com/sersiousSenpai/polis-memory/releases/latest/download/polis-memory-installer.sh | sh
polis init # create ~/.polis and its store
polis hook install # capture what you send in Claude Code
polis mcp install --client claude # let the agent search it
polis doctor # check the install and the chainInstalls the polis binary, then connects Claude Code.