I stopped asking my AI agents to read the project memory. Now the server does it for them.

Description and use

The server pre‑loads project memory into the agent’s first message, ensuring consistent context usage and eliminating the need for agents to remember to search.

Problem

AI agents frequently forgot to consult project memory, leading to inconsistent behavior and silent defects.

Outcome

Injecting pre‑searched memory into the agent’s first message eliminates the need for agents to remember to search, ensuring consistent use of project context and reducing silent defects.

Workflow

• 1. When a session spawns, the server runs a deterministic semantic search on the project memory using a local embedding model.
• 2. The server formats the top‑k results into a structured block prefixed with "=== Project memory (auto‑recalled) ===".
• 3. This block is injected as the first message to the agent, before any user prompt.
• 4. The agent receives the task prompt plus the injected memory block and proceeds to propose a plan.
• 5. If the memory search returns no hits, the injected block is empty and the agent falls back to the original prompt asking it to search.
• 6. End‑to‑end tests seed a test instance with known facts and verify the agent answers correctly without searching the codebase.

Evidence

The server runs the semantic search — before the model gets a single turn — and injects the results directly into the session's first message

Source

A session that reads the code (or the lying README) fails. A session with working recall answers 6443 / callisto / make ship-v2

Source