Read
Agents read from files that are meant to be durable, reviewed and useful across sessions.
Docs
Start with memory agents can trust.
Directive Memory is the first product primitive: durable business context that agents can read, search and update with guardrails.
Directive Memory
What it is.
Most agent workflows fail when context lives across chat history, old files, private notes and half-remembered instructions. Directive Memory gives agents a source of truth they can return to.
It is markdown-native context with MCP tools for reading, searching, patching and refreshing sources. The point is not to make agents sound informed. The point is to make their work grounded, repeatable and reviewable.
How it works
The memory loop.
Search
Hybrid search and graph links help agents find the right context instead of guessing from a single prompt.
Patch
Targeted edits use exact-match or hash-guarded writes so changes are intentional and reviewable.
Receipt
Every meaningful write should leave enough evidence for a human to see what changed and why.
Agent use
What an agent gets.
$ directive memory read clients/acme/ops.md
→ trusted source loaded with content hash
$ directive memory patch clients/acme/ops.md
→ exact edit applied, receipt stored
$ directive memory search "approval workflow"
→ ranked context, not prompt folklore
Current status
Real primitive first, public docs next.
Working now
Memory files, MCP read tools, targeted patching, append-if-missing writes, source refresh and search.
Being improved
Cleaner graph hygiene, better receipt views, stronger source maps and safer cross-workspace conventions.
Not the goal
A chatbot memory trick, a prompt dump, or a black-box knowledge base nobody can inspect.
Next docs
Tasks, Skills and Permissions.
The next public docs should explain how a task becomes claimable work, how a skill turns repeated procedure into reusable operating memory, and how approval rules keep agent actions bounded.