Your agents change. Your project memory shouldn't.
openwolf keeps one project memory across Claude Code, Codex and OpenCode, intercepts the reads and command output that quietly fill your context, and reports what each session actually cost, read from the harness transcript. Pure local file I/O: no API calls, no telemetry, no added latency.
Full lifecycle hooks Claude Code, Codex CLI, OpenCodeContext integration Cursor, Gemini CLI, Antigravity
The cost of a byte is not its size.
It is its size times every call that follows.
The obvious suspects are your prompts and your codebase. Before building 2.x we audited 16 live projects, 6,869 real API calls of transcripts, and neither was the problem. Nearly half of all tool-result tokens arrived through Bash: grep floods, git show dumps, test logs, whole files printed with cat. None of that is in your prompt. All of it is re-read from cache on every later call for the rest of the session.
Coding agents flood their own context with grep dumps and re-read files, then forget your conventions between sessions. OpenWolf governs the Bash channel at the source, keeps context healthy across long sessions, and measures everything it claims from your agent's own transcripts.
One governed grep flood, measured at the rewrite point.
OpenWolf subtracts its own overhead from every figure it reports. Every digest, hint and reminder it injects is counted against the savings it claims.
Everything works invisibly
OpenWolf hooks into your agent's lifecycle. No commands to remember. It just makes every session smarter.
The Bash Output Governor
grep floods, git show dumps, and file re-prints are condensed structurally before they enter context. The full output stays on disk with a pointer. Test failures are never touched. The saving is measured per call, at the rewrite point.
Measured, Verified, Attributed
Real usage from every transcript, per model. Hook delivery verified against the harness's own records. Cache rebuilds attributed to their trigger and cost. If OpenWolf claims it, it can prove it.
Context That Survives
A ~400-token index of your project state at session start. Rules re-surfaced on a cadence to counter within-session decay. Compaction restores the path-scoped rules the platform documents as lost.
Large Repos, Navigated
openwolf find answers location queries in under 1k tokens. openwolf map ranks the important files by personalized PageRank over the import graph. Tree-sitter symbols give exact line ranges for slice reads.
A Committed Team Brain
Conventions, corrections, and bug fixes live in files that travel through git and code review, readable by every agent and every teammate. On Claude Code, they sync both ways with native auto-memory.
Provably Alive
Heartbeats on every hook, a session-start self-test, and install verification on every update. A hook cannot die silently. Plus loopback-only token-auth dashboard, zero shell interpolation, and secrets excluded from every index.
Three steps. Then invisible.
Initialize
Run one command in any project. Creates .wolf/ directory, registers hooks, scans all files.
Work Normally
Start whichever agent you were going to start anyway. Hooks fire invisibly, tracking, learning, enforcing. You never interact with any of it, and switching agents changes nothing.
See the Proof
Every session, OpenWolf learns preferences, logs bugs, and governs output. The dashboard shows tokens verifiably kept out of context, hook health, and what broke your prompt cache.
The .wolf/ directory
Every project gets a .wolf/ folder containing state, learning memory, and configuration. Markdown is the source of truth.
anatomy-index.json
Durable project index: descriptions, token estimates, content hashes, symbols, and the import graph. Rendered to anatomy.md.
cerebrum.md
Learned preferences, conventions, Do-Not-Repeat mistakes. Budgeted, committed, shared across agents and teammates.
STATUS.md
Session handoff. Regenerate it with /handoff; the next session reaches productive context in one small read.
buglog.json
Bug and fix memory with full-text search. The pre-write hook recalls relevant past fixes before an edit repeats them.
hooks/
10 Node.js lifecycle hooks shared by all wired agents. Pure file I/O, no network, no AI calls, heartbeat-monitored.
token-ledger.json
Estimated, measured, and transcript-verified usage per session and agent, plus governor deltas and cache-rebuild attribution.
The enforcement layer
Ten hooks fire on every agent action. Heartbeat-monitored, never permission-bypassing. Pure Node.js. No network, no AI, no extra cost.