A field report · 4 chapters · 6 min
Four chapters
on giving agents
a memory.
How context evaporates, how a graph remembers, how a runtime briefs, and how four agents start sharing one mind.
Context evaporates.
Every session, the same explanation. Every agent, a stranger to your codebase.
You open Claude. You describe your auth flow, your DB schema, your conventions. Token cost: 8,400.
You open Cursor on the same repo. It knows nothing. You explain again. Token cost: 11,200.
You hit Aider. Same explanation. Different LLM, same blank slate. Token cost: 9,700.
You can't remember why you chose JWT over sessions. Neither can your agents. The decision is gone.
You've spent 29,300 tokens rebuilding what you already had. Context didn't compound. It evaporated.
A living map of your codebase.
Files, decisions, conversations. All connected. All remembered.
CodeOrch crawls every file. Builds an AST. Embeds every function. Indexes it across database, graph, and vector stores.
Functions link to files. Files link to PRs. PRs link to decisions. Decisions link back to functions.
Why JWT? Why this schema? Decisions get pinned to the nodes they touch, searchable, queryable, alive.
Every merge updates the graph. Every test adds a node. The graph only gets smarter.
Memory isn't a chat history. It's a structured, queryable graph of everything your team has ever decided.
From task to PR, fully wired.
An agent claims a task. Two seconds later, it knows everything that matters.
A user, an agent, or an issue webhook drops a task into the queue. The queue logs it.
Any MCP-compatible agent claims it: Claude, Cursor, Aider. Claim is atomic. No collisions.
Three parallel queries across graph, database, and vector store collapse into a 2–6k token briefing. ~2 seconds.
Surgical context in, code out. Pre-checks skip files that already exist. Review gate scores 7 metrics.
The decision, the diff, the dependencies, all indexed for the next task. Memory compounds.
Four agents. One brain.
Claude, Cursor, Aider, different tools, same context. One discovery becomes everyone's context.
Every agent reads from the same memory. No tool is a stranger to your codebase.
Claude takes the 2FA task. The graph locks auth/* against collisions.
Cursor takes the rate-limit task. Different files. No conflict. They work in parallel.
Claude finds a bug in the JWT signer. Writes it to the graph. Cursor picks it up automatically.
Both PRs hit the gate. 7 metrics, A–D grade. Pass → ship. The graph remembers everything.
The end of amnesia.
You read it scroll by scroll. Now build it task by task.