
💡imprint turns coding corrections into durable, portable markdown rules that AI agents recall
💡imprint turns coding corrections into durable, portable markdown rules that AI agents recall before every edit. Go CLI + MCP for Cursor and other editors; built-in Shelves indexes project docs; Desk UI for audit. You talk normally—the agent maintains the vault.
·
Conversation becomes durable rules. The agent classifies each turn, links rules to project docs, and recalls before the next edit. You talk normally; you do not maintain the vault.
| When | Example | |
|---|---|---|
| ADD | New long-term preference | Naming rule, workflow, architecture boundary |
| REINFORCE | Same policy again | “Yes, still PascalCase for exports” |
| SUPERSEDE | Policy changed | Narrow scope, widen scope, replace claim |
| IGNORE | Task-only turn | One-off refactor, chit-chat, secrets |
When the user rejects a stored rule (“don’t record that”), the agent finds and calls forget. Write loop: Memory writes.
Vault holds claim, evidence, and optional doc pointers (sources). Shelves indexes markdown under roots so one find returns rules, snippets, and links.
go install github.com/SteamedBread2333/imprint/cmd/imprint@latest
go install github.com/SteamedBread2333/imprint/cmd/imprint-mcp@latest
cd your-project && imprint initMount imprint MCP in your editor (docs/mcp.md). CLI fallback: imprint --json.
| You | imprint | |
|---|---|---|
| 1 | Install + init | Writes imprint.yaml and editor rules (Cursor, Claude Code, Codex, Trae, Workbuddy) |
| 2 | Speak normally | Agent find (vault + shelves) → classify → add / reinforce / … (optional sources) |
| 3 | Browser audit (optional) | imprint up → imprint desk open |
flowchart TB
U((User<br/>speaks normally))
subgraph Agent["Agent · imprint-mcp"]
direction TB
F["find(scope, query)"]
C{"Classify<br/>ADD · REINFORCE · SUPERSEDE · IGNORE"}
W["add / supersede<br/>optional sources"]
R["find again before coding · cite [r-id]"]
end
subgraph Store["imprint storage"]
direction LR
Vault[("~/.imprint/projects/…/vault.db<br/>claim · evidence · sources")]
Shelves[("…/state/shelves.db<br/>roots doc BM25")]
end
subgraph FindOut["one find call"]
direction LR
FR["rules<br/>resolved_sources"]
FD["documents<br/>snippet"]
FL["links"]
end
subgraph Audit["Optional · human audit"]
H((You)) --> Desk["desk · show"]
end
UP["imprint up"] -.->|host indexes| Shelves
U -->|speak / task| Agent
F --> Vault
F --> Shelves
Vault --> FR
Shelves --> FD
Vault --> FL
Shelves --> FL
FR & FD & FL --> C
C -->|persist| W
W --> Vault
R --> F
Desk --> Vault
Desk --> ShelvessequenceDiagram
participant U as User
participant A as Agent
participant V as vault
participant S as shelves
U->>A: new task / continue coding
A->>V: find(scope, query)
V->>S: BM25 docs (when shelves on)
V-->>A: rules · documents · links
A->>A: code from imprints + excerpts
Note over A: cite [r-id] and doc pathCommit imprint.yaml (roots, plugin switches, tunables). Private runtime lives under ~/.imprint/projects/<hash>/ (hash of the repo absolute path). Nothing under that tree belongs in git.
Default vault: walk up for imprint.yaml, then ~/.imprint/projects/<hash>/. Override with vault: in yaml, --vault, or IMPRINT_VAULT. --global uses the same layout keyed off cwd when no project is found.
imprint.yaml # git: host + shelves.roots + plugins + supersede.inheritance_alpha
docs/ # typical shelves root
.cursor/rules/ # typical shelves root
~/.imprint/projects/<hash>/ # private runtime (outside the repo)
project.root # links back to this repo (for sources / MCP)
vault.db # SQLite vault (rules, evidence, edges, sources)
state/
shelves.db # rebuildable doc index
telemetry/ # optional metrics
plugins/ # plugin derived state
export/ # optional json/jsonl projectionsVault and shelves open SQLite in WAL mode so CLI, MCP, and host can read and write across processes. -wal / -shm next to a *.db are normal; do not delete them while imprint is running.
Rules live in vault.db. IDs: r-YYYY-MM-DD-NNN. Status: active | dormant | superseded. supersede marks old rules superseded; sweep decays stale rules to dormant; forget deletes. Interactive graph: imprint desk open (host GET /graph).
| Where it runs | Config | |
|---|---|---|
| Shelves | host | shelves.enabled, config.roots in imprint.yaml |
| Desk | External plugin | plugins.desk + imprint-desk-plugin |
| Embed | External sidecar (plugin protocol) | ~/.imprint/embed.yaml + imprint-embed-sidecar |
Embed adds a semantic duplicate gate to add: the lexical Jaccard check only catches near-identical wording, so paraphrases slip through. Off by default; when the sidecar is absent or slow, writes degrade back to the lexical path. imprint-mcp starts the sidecar if the port is down (reuses it if healthy). imprint plugin start embed always stop-then-fork; imprint plugin stop embed stops it. imprint up / down do not touch embed. See docs/semantic-dedup.md for the full pipeline, calibration data, and tuning.
Shelves indexes markdown under roots (e.g. docs/, .cursor/rules/). One MCP mount (imprint-mcp); with shelves on, find + query returns rules, documents, and links. See docs/shelves-builtin.md.
Global flags: --vault PATH · --global · --json (machine-readable stdout for agents)
| Command | What it does |
|---|---|
imprint init | Write imprint.yaml (if missing) and editor agent rules |
imprint find [--scope a,b] [--query TEXT] | Recall imprints (scope AND); with query + shelves → also documents, links |
imprint add CLAIM --scope a,b --text ORIG | Create an imprint (--sources JSON or MCP sources links docs) |
imprint reinforce ID --evidence TEXT | Strengthen a rule after explicit reaffirmation (diminishing gain, cap 0.95) |
imprint supersede OLD --claim NEW --scope a,b | Replace a rule; old → superseded |
imprint forget ID | Delete permanently (lifecycle event remains) |
imprint list · show · get ID · report | Browse, inspect, and audit lifecycle / recall / telemetry |
imprint sweep · export | Decay stale rules · write <vault>/export/vault.json or vault.jsonl (--format jsonl) |
imprint find --scope go,naming --query PascalCase
imprint add "Use snake_case" --scope python,naming --text "user said snake_case"
imprint get r-2026-09-11-001| Command | What it does |
|---|---|
imprint up | Start the local stack (vault API, shelves, enabled desk, …). Does not start or stop embed |
imprint down | Stop the local stack. Leaves embed running (imprint plugin stop embed) |
imprint desk open | Open desk in the browser (requires up first) |
imprint status | Snapshot of running services |
imprint up && imprint desk open
# when done:
imprint downAfter editing imprint.yaml: run down then up.
| Command | What it does |
|---|---|
imprint host start / host stop | Host only (includes shelves) |
imprint plugin start ID / plugin stop ID | External plugins only (id required; pass embed to start/stop the sidecar) |
plugin list · enable · disable | Toggle plugins in yaml |
imprint host serve [--listen ADDR] | Foreground host (Ctrl+C) |
imprint clear --confirm --yes | Delete every rule — irreversible |
imprint version | Print version |
Run imprint --help or imprint help <cmd> for full flags.
find returns claim/count metadata and bounded document snippets by default. Document hits use shelves.config.find_top_k (default 2, one chunk per path), not rule top_k; snippets are a query-window of snippet_runes (default 80). get r-… folds evidence and source text. Use include_evidence or full only for audit. Dormant rules can contribute at most one penalized wake_candidate; only an explicit reinforce wakes one.add, reinforce, and supersede apply field-aware privacy checks (stricter on text/evidence; claim still blocks secrets, PEM, and government IDs). Real tokens or PEM in any field still reject. Source paths must remain inside the workspace and may not target credentials, .env*, *.pem, or *.key. add also rejects high-similarity active duplicates. reinforce requires non-empty evidence and uses diminishing confidence gain; find hits only update recall stats.supersede.inheritance_alpha in imprint.yaml (default 0.20; 0 copies old, 1 drops to baseline).ts, tsx, golang) canonicalize via GitHub Linguist (go-enry); non-language tags pass through.imprint report --days 30 summarizes lifecycle events, duplicates, conflicts, zero-recall rules, and telemetry latency. Telemetry is a daily JSONL under <vault>/state/telemetry/ and never stores query, claim, evidence, or path text.sources point rules at doc paths or headings; compact get r-… returns pointers, full:true resolves excerpts. Design: docs/imprint-shelves-linking.md.| Topic | English | 中文 |
|---|---|---|
| MCP mount | docs/mcp.md | docs/mcp.zh.md |
Editor init | docs/editors.md | docs/editors.zh.md |
| Shelves & linking | docs/shelves-builtin.md · docs/imprint-shelves-linking.md | docs/shelves-builtin.zh.md · docs/imprint-shelves-linking.zh.md |
| Write loop & scenarios | docs/correction.md | docs/correction.zh.md |
| Testing & acceptance | docs/testing.md | docs/testing.zh.md |
MCP mount examples: docs/mcp.md · docs/mcp.zh.md
docker pull ghcr.io/steamedbread2333/imprint:latest # or GitHub Releases binaries
make install # from clone
make publish V=X.Y.Z # tag + CI → Release + GHCRLocal builds without a tag print devel.
import "github.com/SteamedBread2333/imprint/pkg/imprint"
dir, _ := imprint.ResolveDir("", false, nil, nil, nil)
v, _ := imprint.Open(imprint.OpenOptions{Dir: dir})
v.Add("Use gofmt", []string{"go"}, "gofmt", 0.6)
v.Find([]string{"go"}, "", 5)MIT
Paste this prompt into your agent. It reads this page and does the setup for you.
Read https://aiagentslisting.com/mcp/imprint to learn what the "imprint" MCP server does and how to install it. Add it to my coding agent's MCP configuration as documented on that page, then confirm the server connects and list the tools it exposes.Agents can also browse this directory over MCP at https://aiagentslisting.com/api/mcp
go install github.com/SteamedBread2333/imprint/cmd/imprint@latest
go install github.com/SteamedBread2333/imprint/cmd/imprint-mcp@latestThis server doesn't publish a machine-readable tool manifest, and it doesn't expose a public endpoint we could ask. Connect the server locally to see what it exposes.
We could not run this server: it ships no npx package for our sandbox and no public endpoint. The four checks that need a live server are left out, and the grade tops out at B until one exists.
Maintain this server? Claiming proves you own the listing and earns the last 4 points.
Claim this listingPaste this line near the top of your repository's README. The badge always renders the current score.
[](https://aiagentslisting.com/mcp/imprint)Nothing comparable is listed yet.
v1.3.14imprint v1.3.14Release notesv1.3.13imprint v1.3.13Release notesv1.3.12imprint v1.3.12Release notesv1.3.11imprint v1.3.11Release notesv1.3.9imprint v1.3.9Release notesv1.3.8imprint v1.3.8Release notesv1.3.7imprint v1.3.7Release notesv1.3.6imprint v1.3.6Release notesv1.3.5imprint v1.3.5Release notesv1.3.4imprint v1.3.4Release notesv1.3.3imprint v1.3.3Release notesv1.3.2imprint v1.3.2Release notesv1.3.1imprint v1.3.1Release notesv1.2.1imprint v1.2.1Release notesv1.2.0imprint v1.2.0Release notesv1.1.2imprint v1.1.2Release notesv1.1.1imprint v1.1.1Release notesv1.1.0imprint v1.1.0Release notesv1.0.0imprint v1.0.0Release notesConnect to the AI Agents Listing MCP server at https://aiagentslisting.com/api/mcp using the HTTP transport and complete the sign-in when it asks. Then call the submit_review tool for the listing "imprint" with a rating from 1 to 5 and a short note on what worked and what did not.No reviews yet
Be the first to share your experience with this tool.
One email a week. New agents, MCP servers and skills, and what is actually getting traction.