StateJar memory

StateJar integrations: memory your AI agents can trust

statejar.com · Docs · Blog · Pricing · LinkedIn

Client code that connects AI coding agents and assistants to StateJar, a memory layer that keeps facts instead of transcripts. Each value is checked by its type (a value that fails is refused, with a reason), every change is a new version with the old one kept in history, an assistant gets only the facts a question needs, every disclosure is audited, and each state has a verifiable handle (shm_…, the SHA-256 of the canonical state).

Quick start

1. Make a developer key at https://statejar.com/api-keys and set STATEJAR_API_KEY.
2. Pick your client:

Details for each are below.

What is here

• Hosted MCP server, for any client that speaks MCP over HTTP: https://statejar.com/mcp (live).
• Claude Code plugin: plugins/statejar/; guide in claude-code/. Marketplace: this repository.
• Codex plugin (the same plugin): guide in codex/. Marketplace: this repository.
• OpenCode plugin: opencode/. npm: opencode-statejar.
• Hermes Agent memory provider (memory.provider: statejar): hermes/statejar/. Install with hermes plugins install from this repository.
• OpenClaw memory-slot plugin: openclaw/. npm: openclaw-statejar.
• Agent skill, for any agent that reads SKILL.md: skills/statejar/.
• CLI: pip install statejar gives statejar login, init, add, search, list, forget, export, status.

Every integration uses a StateJar developer key (https://statejar.com/api-keys), scoped to your account. None of them stores a transcript. Each removes secrets before anything is sent.

Availability. Recall and capture call StateJar's REST API at https://api.statejar.com. The plugins' StateJar tools use the hosted MCP server at https://statejar.com/mcp. Both are live.

Any MCP client

Make a developer key at https://statejar.com/api-keys and put it in STATEJAR_API_KEY. Then:

# Claude Code

claude mcp add --transport http statejar \ https://statejar.com/mcp \ --header "Authorization: Bearer $STATEJAR_API_KEY"

Codex

codex mcp add statejar \ --url https://statejar.com/mcp \ --bearer-token-env-var STATEJAR_API_KEY

OpenCode, in opencode.json:

{

"mcp": { "statejar": { "type": "remote", "url": "https://statejar.com/mcp", "oauth": false, "headers": { "Authorization": "Bearer {env:STATEJAR_API_KEY}" } } } }

Any other client: a remote (HTTP) MCP server at https://statejar.com/mcp with the header Authorization: Bearer . The ten tools are remember, search, get, list, update, delete, history, audit, verify and export.

Claude Code

claude plugin marketplace add KING-OF-FLAME/statejar-integrations

claude plugin install statejar@statejar

See claude-code/README.md.

Codex

codex plugin marketplace add KING-OF-FLAME/statejar-integrations

codex plugin add statejar@statejar

Then trust the plugin's hooks with /hooks. See codex/README.md.

OpenCode

Add "plugin": ["opencode-statejar"] and StateJar's MCP server to opencode.json. See opencode/README.md.

Hermes Agent

hermes plugins install KING-OF-FLAME/statejar-integrations/hermes/statejar --enable

hermes memory setup # choose statejar

See hermes/statejar/README.md.

OpenClaw

openclaw plugins install openclaw-statejar

openclaw statejar init # the memory slot, autoRecall, autoCapture

See openclaw/README.md.

Agent skill

Copy skills/statejar/ into your agent's skills folder (for Claude Code, ~/.claude/skills/). See skills/README.md.

Measured results

Numbers from the StateJar service these clients talk to, measured by StateJar's own evaluation jobs; intervals are 95% (Wilson). These clients add no extraction of their own.

• Release gate, extraction (2026-11-06, gpt-6-luna candidate, 345-message dev subset): precision 423/441 = 95.9% [93.6–97.4], recall 423/467 = 90.6% [87.6–92.9].
• Real-world messages (2026-11-05, gpt-6-luna, 300 messages, rw3): precision 232/247 = 93.9% [90.2–96.3], recall 232/276 = 84.1% [79.3–87.9]; $0.54 per 1,000 adds; p95 13.1 s per message.
• Real-world messages (2026-11-05, gemini-3.5-flash-lite paid, same 300): precision 243/262 = 92.8% [88.9–95.3], recall 243/276 = 88.0% [83.7–91.4].
• Wrong facts in answers (2026-11-03, gemini-3.5-flash-lite, 120 conversations): 8/120 = 6.7% [3.4–12.6] with StateJar; 1/120 = 0.8% with the full transcript, which did better on these short conversations.

The full tables, with the cost per 1,000 adds of each model, the tokens per turn by chat length and which model production uses (Gemini gemini-3.5-flash-lite first, then Mistral, then OpenRouter; gpt-6-luna is the staging candidate), are in the StateJar service's README.

Tests

python -m pytest tests/

claude plugin validate ./plugins/statejar claude plugin validate .

End to end, against a local or staging StateJar (never production): a fact stated in one session of the real client is recalled in a fresh one, through the hooks and through StateJar's MCP tools. Each run makes a throwaway account and deletes it. With no model on the machine, a scripted stand-in (e2e/scripted_model.py) answers "OK" or calls StateJar's search and repeats its output verbatim; codex_e2e.py --model ollama: uses a real local model.

Each script takes the API (and, where it uses MCP, the MCP URL) of a local or staging server, and an output folder:

API=http://127.0.0.1:8000/api/v1

MCP=http://127.0.0.1:8000/mcp

python e2e/codex_e2e.py --api $API --mcp $MCP --out /tmp/codex-e2e python e2e/claude_code_e2e.py --api $API --mcp $MCP --out /tmp/cc-e2e python e2e/opencode_e2e.py --api $API --mcp $MCP --out /tmp/oc-e2e python e2e/hermes_e2e.py --api $API --out /tmp/hermes-e2e python e2e/openclaw_e2e.py --api $API --out /tmp/openclaw-e2e

• openclaw_e2e.py needs Node 24.16+.
• python e2e/openclaw_discovery.py and python e2e/codex_discovery.py show what OpenClaw and Codex themselves load from the plugin (no model, no StateJar).
• HERMES_PYTHON=/path/to/hermes/venv/bin/python python -m pytest tests/test_hermes_provider.py runs the provider test with Hermes's own loader.

Licence

MIT (see LICENSE). The StateJar service these clients talk to has its own terms: https://statejar.com/terms.