@ghostpaw/codex

Agent memory as beliefs with a source, a certainty and a revision history.

ghostpaw-codex --help

Local, evidence-backed memory for agents. One SQLite file, zero runtime dependencies.

Codex stores what an agent believes as claims with a source, a certainty, evidence counts, decay and an honest revision history. Agents search before they remember, confirm or correct instead of appending, and see which beliefs are stale, fading or weakly evidenced. It runs as a TypeScript library, an MCP server, a CLI and an Agent Skill, all generated from the same seven tools.

Website: codex.ghostpaw.dev

The release binary includes its runtime. npm and library use require Node.js 24+.

Install

Release binary (default)

Each line downloads the latest release asset, checks it against the release's SHA256SUMS and installs ghostpaw-codex into ~/.local/bin. On Linux and macOS the trailing export puts that folder on PATH for the current shell only: neither the line nor install.sh edits ~/.profile or any other shell startup file. The Windows script adds its folder to the user PATH (open a new terminal afterwards).

linux-x64:

curl -fsSL https://raw.githubusercontent.com/GhostPawJS/codex/main/install.sh | sh -s -- --target linux-x64 && export PATH="$HOME/.local/bin:$PATH"

linux-arm64:

curl -fsSL https://raw.githubusercontent.com/GhostPawJS/codex/main/install.sh | sh -s -- --target linux-arm64 && export PATH="$HOME/.local/bin:$PATH"

darwin-arm64:

curl -fsSL https://raw.githubusercontent.com/GhostPawJS/codex/main/install.sh | sh -s -- --target darwin-arm64 && export PATH="$HOME/.local/bin:$PATH"

darwin-x64:

curl -fsSL https://raw.githubusercontent.com/GhostPawJS/codex/main/install.sh | sh -s -- --target darwin-x64 && export PATH="$HOME/.local/bin:$PATH"

windows-x64 (PowerShell):

irm https://raw.githubusercontent.com/GhostPawJS/codex/main/install.ps1 | iex

To keep ~/.local/bin on PATH in new shells, append it to ~/.profile once (this is the only step that edits a startup file):

echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.profile

To pin a release instead of taking the latest, pass --version:

curl -fsSL https://raw.githubusercontent.com/GhostPawJS/codex/main/install.sh | sh -s -- --target linux-x64 --version 0.6.0 && export PATH="$HOME/.local/bin:$PATH"

$ ghostpaw-codex --version
ghostpaw-codex 0.6.0 (binary, linux-x64)
$ ghostpaw-codex --version --short
0.6.0
$ ghostpaw-codex --version --json
{"name":"ghostpaw-codex","version":"0.6.0","variant":"binary","target":"linux-x64"}

variant is binary, npm or library; target is null outside the release binary.

A session with the 0.5.2 linux-x64 binary against a small database of five beliefs:

The same session as an animated GIF.

Other release installers:

mise use -g github:GhostPawJS/codex
ubi --project GhostPawJS/codex --exe ghostpaw-codex --in ~/.local/bin
eget GhostPawJS/codex

For aqua, define a github_release package in aqua.yaml:

registries:
  - name: codex
    type: local
    path: codex-registry.yaml
packages:
  - name: GhostPawJS/codex@v0.6.0
    registry: codex

codex-registry.yaml:

packages:
  - type: github_release
    repo_owner: GhostPawJS
    repo_name: codex
    asset: ghostpaw-codex-{{trimV .Version}}-{{.OS}}-{{.Arch}}.tar.gz
    format: tar.gz
    replacements: { amd64: x64 }
    overrides:
      - goos: windows
        asset: ghostpaw-codex-{{trimV .Version}}-windows-{{.Arch}}.zip
        format: zip
    files:
      - name: ghostpaw-codex

Install scripts verify SHA256SUMS. To verify a downloaded archive's build attestation:

gh attestation verify <archive> -R GhostPawJS/codex

Set GHOSTPAW_VERIFY=attestation for attestation verification in install.sh. Binaries are unsigned; macOS binaries carry an ad-hoc signature. For a browser download on macOS, remove quarantine with xattr -d com.apple.quarantine /path/to/ghostpaw-codex. Linux binaries require glibc 2.28+; use npx on Alpine. See the binary build workflow.

Approximate sizes for 0.5.2 (Node 26.12.1 inside each binary):

Target	Archive	Binary
linux-x64	41.7 MB tar.gz	117.4 MB
linux-arm64	41.2 MB tar.gz	115.2 MB
darwin-arm64	44.8 MB tar.gz	146.7 MB
darwin-x64	46.4 MB tar.gz	149.5 MB
windows-x64	38.4 MB zip	105.6 MB

The npm package is about 920 KB unpacked and needs Node 24+. Every release is installed and run on ubuntu-24.04, macos-15 and windows-2025 by the install check.

npx / npm (for npm-based environments)

The npm package requires Node 24+:

npm install -g @ghostpaw/codex
npx -y @ghostpaw/codex --help

Library (for harnesses embedding it in code)

Requires Node 24+:

npm install @ghostpaw/codex

import { initCodexTables, read, write } from '@ghostpaw/codex';

See Library usage.

30-second quickstart

MCP server

Every client below starts the same stdio server: ghostpaw-codex mcp. Use the full executable path if its folder is not on PATH. For npm-based environments, use npx -y @ghostpaw/codex mcp. Beliefs live in ~/.ghostpaw/codex.db unless the GHOSTPAW_CODEX_DB environment variable (absolute path) or --db <path> says otherwise.

Claude Code

claude mcp add ghostpaw-codex -- ghostpaw-codex mcp

or as a plugin, which also installs the skill:

/plugin marketplace add GhostPawJS/codex
/plugin install ghostpaw-codex@ghostpaw-codex

Codex CLI (~/.codex/config.toml)

[mcp_servers.ghostpaw-codex]
command = "ghostpaw-codex"
args = ["mcp"]

Cursor (.cursor/mcp.json), Claude Desktop (claude_desktop_config.json) or a project .mcp.json:

{
  "mcpServers": {
    "ghostpaw-codex": {
      "command": "ghostpaw-codex",
      "args": ["mcp"]
    }
  }
}

Cline (MCP Servers → Configure MCP Servers, cline_mcp_settings.json). VS Code often starts without ~/.local/bin on PATH, so use the absolute path that echo "$HOME/.local/bin/ghostpaw-codex" prints (PowerShell: "$env:LOCALAPPDATA\Programs\ghostpaw\bin\ghostpaw-codex.exe"). Merge the entry into the existing mcpServers. Settings file locations, the cline mcp add form and checks: llms-install.md. For Cline CLI 3.x, cline config mcp lists the configured servers.

{
  "mcpServers": {
    "ghostpaw-codex": {
      "command": "/absolute/path/to/ghostpaw-codex",
      "args": ["mcp"],
      "env": { "GHOSTPAW_CODEX_DB": "/absolute/path/to/.ghostpaw/codex.db" }
    }
  }
}

Claude Desktop users can also install server.mcpb from the latest GitHub release. A Dockerfile is included for container hosts: docker build -t ghostpaw-codex . && docker run -i --rm -v codex:/data ghostpaw-codex.

The server speaks MCP 2026-07-28 (stateless, server/discover) and the initialize-based revisions 2025-11-25, 2025-06-18, 2025-03-26 and 2024-11-05.

CLI

ghostpaw-codex search <query> [--limit N --min-score N --category C --source S]
ghostpaw-codex review <flags|status|log|trends> [--limit N --now N]
ghostpaw-codex belief add <claim> --source S --category C [--certainty N --provenance P]
ghostpaw-codex belief show <beliefId>
ghostpaw-codex belief retire <beliefId>
ghostpaw-codex belief confirm|correct|forget|defer <beliefId> [flags]
ghostpaw-codex belief merge <beliefIds...> [flags]
ghostpaw-codex proximity dismiss <beliefA> <beliefB>
ghostpaw-codex mcp|skill|tools|describe|call|completions|version|help

Belief additions include a nearest-belief hint when an active belief has similarity of at least 0.7; use it to confirm or merge related claims.

WARNING: belief retire hard-deletes the whole lineage (the belief and all its predecessors/successors) and cannot be undone. Use belief forget for soft removal that keeps lineage. Every retire invocation warns on stderr.

--json is accepted on every read command, including all review views, belief show, alias list, tools, describe, version, help and skill guides. Set GHOSTPAW_OUTPUT=json to opt all commands into JSON output as if --json were passed. Other values or an unset variable preserve default output, including when stdout is piped or a TTY.

Global flags can appear anywhere: --db <path>, --json, --quiet, --compact, --help/-h, --version/-V/-v, --no-color. NO_COLOR is accepted; output uses no terminal colors. Noun-verb commands print short text by default, including when stdout is piped. Search and review lists show one belief per line: id claim [certainty source strength]. Flags append reason codes; the log shows id type claim. Add, show, correct and merge show aligned detail fields and nearby beliefs. Status and trends show aligned key: value lines. --json prints the full tool result and --compact removes JSON indentation. Legacy tool invocation keeps its JSON output. --quiet prints only result ids for search, review flags and log, and the affected belief id for writes; status, trends and version print nothing.

Use help belief, belief --help or belief correct --help for schema-derived help. Each revision verb accepts only the flags it uses. Correction takes --claim; forgetting also accepts --successor-id. Deferral accepts --deferred-until as YYYY-MM-DD, DD.MM.YYYY, today, tomorrow, +3d, an ISO date-time with timezone, or epoch milliseconds. Calendar dates resolve to local end of day and must be in the future. Array flags repeat, kebab-case flags map to schema properties, and values are literal, including leading @. Use JSON stdin (-) for long input. -- ends flags. A positional -, or piped stdin with no arguments, reads a JSON object; explicit flags override its fields. Database selection uses --db, GHOSTPAW_CODEX_DB, then ~/.ghostpaw/codex.db. Opening the default database prints using default db <path> once on stderr. Suppress the notice with --quiet or GHOSTPAW_QUIET=1. Commands that do not open a database print no notice. Top-level help shows the resolved path. Version output starts with ghostpaw-codex and identifies binary target, npm or library execution.

ghostpaw-codex skill prints the full skill reference. Use ghostpaw-codex skill --list to list workflow guides and ghostpaw-codex skill capture-beliefs-well to read a guide. skill --help shows these forms; unknown guide names include a suggestion.

Exit	Noun-verb result
0	Success, including no_op
1	Tool error
2	Usage or schema error
3	Clarification needed or not found
4	Conflict or version mismatch
70	Internal error

Errors and recovery hints go to stderr. Clarification prints the question and missing flags to stderr. With --json, usage and internal errors also print {ok:false,error:{code,message}} to stdout. Tool results retain their full shape. Legacy tools [--json], describe <tool>, call <tool> [JSON|-] and <tool_name> --parameter value keep their existing exit codes. completions bash, completions zsh and completions fish print shell scripts.

ghostpaw-codex belief add "Deploys go out on Tuesdays" --source explicit --category procedure

ghostpaw-codex search "when do we deploy" --json | jq '.data.results'

printf '%s' '{"beliefId":1}' | ghostpaw-codex belief show - --json

File values and leading @

File input is opt-in for each CLI argument. When an argument's help lists (@file, @- stdin, @@ literal @), @path reads a UTF-8 file, @- reads stdin, and @@text passes the literal value @text.

No current Codex argument enables this option: flags and positional values keep leading @ and @@ as written. For long claims, pass a JSON object through stdin with - (see the example above).

Agent Skill

SKILL.md (repo root) and skills/codex/SKILL.md follow the Agent Skills format and ship in the npm package. Copy skills/codex/ into your agent's skills folder, or print it:

ghostpaw-codex skill > SKILL.md

Library

npm install @ghostpaw/codex

import { DatabaseSync } from 'node:sqlite';
import { initCodexTables, read, write } from '@ghostpaw/codex';

const db = new DatabaseSync('codex.db');
initCodexTables(db);

const { id } = write.remember(db, {
  claim: 'The API uses GraphQL with a REST fallback.',
  source: 'explicit',
  category: 'fact',
});
write.confirmBelief(db, id);
write.correctBelief(db, id, {
  claim: 'The API uses GraphQL exclusively; REST was retired.',
  source: 'observed',
  category: 'fact',
});

read.recall(db, 'graphql api'); // ranked, lineage-aware
read.listFlags(db);             // stale, fading, low-trust, single-evidence

Wire the same tools into any LLM SDK. Each tool carries a JSON Schema:

import { tools } from '@ghostpaw/codex';

for (const tool of tools.codexTools) {
  // tool.name, tool.description, tool.inputSchema, tool.handler(db, input)
}

Or embed the MCP server in your own process:

import { createCodexMcpServer } from '@ghostpaw/codex/mcp';
const server = createCodexMcpServer({ dbPath: './codex.db' });

Docs

The live demo runs the library in your browser against an in-memory SQLite database. Each view below opens a different part of it:

• Dashboard: integrity score and recent beliefs
• Capture: record a belief and see close matches you already hold
• Flags: beliefs flagged for review: confirm, correct or forget
• Activity Log: activity log of every write

Why not MEMORY.md?

A markdown memory file works until it doesn't:

• It only grows. Agents append, rarely delete, and the file ends up in every prompt. Codex returns the few beliefs that match a query (FTS5 bm25 plus hashed character-trigram similarity, IDF-weighted with a 55% relative cutoff).
• Contradictions pile up. "We use REST" and "REST was retired" sit side by side. Codex detects near-duplicates on every write, and corrections supersede the old belief while keeping the lineage.
• Everything looks equally true. A guess and a confirmed fact read the same in prose. Codex tracks source (explicit, observed, distilled, inferred), certainty and evidence count, and flags what is stale or weakly supported.
• No time. Codex decays freshness, so old beliefs surface for review instead of quietly steering behaviour.
• Concurrent edits. Two agents writing one file clobber each other. Codex is a SQLite database in WAL mode.

It stays local: one file you can copy, back up or inspect with any SQLite tool. No server, no API key, no network.

Tools

Tool	What it does
search_codex	Lexical recall with FTS5 bm25 and character-trigram similarity
review_codex	Dashboards: flags, status, trends, log
inspect_codex_item	One belief in full: metrics, lineage, nearby beliefs
remember_belief	Capture a claim with source, category and provenance
revise_belief	Confirm, correct, forget, defer or merge
retire_belief	Hard-delete for privacy or cleanup
dismiss_proximity	Mark two nearby beliefs as intentionally distinct

Results are structured (outcome: success | no_op | needs_clarification | error) with entities, next-step hints and recovery advice. The MCP server, CLI, SKILL.md and plugin manifests are generated from this registry (npm run surfaces) and drift-checked in CI.

Performance

Measured with npm run bench on Node 24.14.0, Linux, Intel Xeon (shared cloud vCPU), 10,000 beliefs in a file-backed database:

Operation	p50	p95
search_codex (tool)	3.6 ms	5.0 ms
MCP tools/call search, stdio framing included	3.8 ms	4.2 ms
read.recall	3.7 ms	4.8 ms
inspect_codex_item	3.5 ms	3.8 ms
write.remember (includes proximity scan)	4.5 ms	5.6 ms
review_codex	11.7 ms	13.2 ms

Cold start (process spawn to first response, median): --version 26 ms, MCP server/discover 24 ms, first search against the 10,000-belief file 53 ms. Seeding 10,000 beliefs takes 2.2 s on disk.

Gzip sizes and cold-start medians are budgeted in budgets.json and enforced by npm run check.

The model

Concept	Purpose
Belief	A claim held with certainty, evidence, provenance, decay and revision lineage
Dismissal	A record that two beliefs close by character-trigram vectors are intentionally distinct
Flag	A derived attention signal: stale, fading, low-trust, single-evidence or gap
Proximity	Lexical nearness between beliefs (character-trigram vectors), detected on write
Lineage	The revision chain from corrections and supersessions, with version diffs

State is derived, never toggled: freshness, strength, review priority and flags are computed at read time.

Package surface

Import	Contents
@ghostpaw/codex	initCodexTables, read, write, tools, skills, soul, all types
@ghostpaw/codex/mcp	createCodexMcpServer, serveCodexMcp, generic MCP server helpers
ghostpaw-codex (bin)	Noun-verb CLI, legacy tool forms, mcp, skill, shell completions
dist/mcp-server.js	Standalone stdio MCP server entry (used by the MCP Bundle and the Dockerfile)
SKILL.md, skills/codex/	Agent Skill with 17 workflow references

ESM and CommonJS, with type declarations for both.

Documentation

Document	Audience
HUMAN.md	Developers using the read / write API
LLM.md	Agent builders wiring soul, tools and skills
docs/README.md	Architecture: model, invariants, scoring, source layout
docs/entities/	Per-entity manuals with exact public API listings
CHANGELOG.md	Release history

Development

npm install
npm run check   # types, lint, zero-deps, tests, build, surfaces, size, API, consumer, startup, ratchet
npm run bench   # hot-path benchmarks at 10,000 beliefs
npm run surfaces  # regenerate SKILL.md, plugin manifests, server.json, mcpb manifest

npm run check is the single gate, locally and in CI. Size budgets (budgets.json) and the public API snapshot (api-surface.json) can only grow with an explicit approval tag in the commit message. The repo is pinned to Node 24.14.0 via .nvmrc.

Support

If this package helps your project, consider sponsoring its maintenance:

---

Anonyfox • MIT License

Query aliases

Search is lexical: FTS5 bm25 plus hashed character-trigram similarity, with IDF-like token weighting and a 55% relative cutoff. It does not provide semantic understanding. In corpora of at least 20 active beliefs, queries matching only common terms return at most three results with a partial_match warning.

Aliases are off by default (the table starts empty). Add directed, one-hop aliases when query vocabulary differs from stored claims:

ghostpaw-codex alias add class warrior
ghostpaw-codex alias list
ghostpaw-codex alias remove class warrior

The library equivalents are write.addQueryAlias(db, term, alias, options?), write.removeQueryAlias(db, term, alias) (returns whether removed), and read.listQueryAliases(db) (sorted { term, alias, createdAt } records). Both words are normalized with normalizeClaim and must be distinct single words. Adding an existing pair preserves its creation time; options.now sets the time on first insertion. Expansion is deduplicated and never follows alias chains; the query vector uses the original query.

Provenance and revision compatibility

Belief records include provenanceSources, an append-only list of distinct origin references. Every write with --provenance adds a reference, including duplicate adds and unchanged corrections. Corrections and merges carry all sources forward. The legacy provenance field remains readable with its existing single-value behavior; use provenanceSources for the complete history.

Superseded-belief errors and belief show identify the current active successor. Library details also expose activeSuccessorId, while supersededBy retains the immediate lineage link. Correcting to the same normalized claim and metadata returns no_op. A merge into --successor-id can omit --claim.

belief --help lists belief verbs. review flags --now +8d and review flags --now 2026-11-10 accept the same date formats as deferral, including historical dates and epoch milliseconds. Relative day offsets also accept -3d and 30d. Calendar dates and relative days use local end of day.

New CLI and MCP database files use mode 0600; newly created parent directories use 0700. Existing permissions are preserved. On first upgrade, legacy active normalized duplicates consolidate into the oldest id, preserving summed evidence, source references and lineage. Later --allow-duplicate additions remain separate. Merge fan-in alone does not trigger unstable; repeated corrections still do.

Search treats punctuation literally; wrap the whole query in double quotes for a phrase. Use search --fts 'alpha OR beta' for raw FTS5 syntax, or search note --prefix "project:" to filter claims by a case-insensitive prefix after leading whitespace. A query is required with --prefix. Library recall options and search tool inputs also accept fts and prefix. Default search ranks claims whose entity before the first colon matches the leading query tokens first, ignoring case and treating spaces and hyphens alike (for example, Agentpreneur: ... for Agentpreneur state files routine memory db). This also applies with --prefix; --fts and quoted phrase ranking remain unchanged.

Review status integrity remains the percentage of active beliefs with certainty at least 0.7. The new, unconfirmed count covers beliefs with one evidence item younger than seven days; it overlaps the existing strength counts.

Review: review status --by-prefix counts active claim prefixes; review duplicates --limit 20 lists nearby active pairs immediately, honoring dismissals.