Project memory for a coding agent

Problem: A coding agent forgets the project's decisions, the user's corrections and its rules in every new session; it repeats the same mistakes and asks the same questions.

Solution: Connect the agent to Agent Memory over MCP. Have it recall at the start of a task, write what is durable during the task, distill at the end, and report the outcome. Put fixed project-wide information in blocks.

1. ConnectCopy Anchor Link

In the console, create an API key for your project (read, write, feedback, curate). Add it to your client:

{
  "mcpServers": {
    "agent-memory": {
      "url": "https://memory.empecta.com/v1/mcp",
      "headers": { "Authorization": "Bearer amk_..." }
    }
  }
}

Keyless sign-in through the browser (OAuth) is on the roadmap; for now every client connects with an API key. Details: Connect with MCP.

2. Write the project's fixed information to blocksCopy Anchor Link

For short information that applies to every task:

curate_block(label="tech_stack", action="add", value=".NET 10, PostgreSQL 17, Redis 7, EF Core")
curate_block(label="coding_conventions", action="add", value="All entity IDs are Guids; async methods end with the Async suffix")

Blocks appear at the top of every search_memory output, regardless of the query.

3. The agent's loopCopy Anchor Link

The server passes this loop to the agent as instructions; also add it to your client's project instructions (for the text, see Connect with MCP).

At the start of a task:

search_memory(query="numbering and error format in the invoicing module", runId="issue-482")

The output is split into sections: blocks, platform/tenant rules, workspace history, procedures, similar tasks and this run's current memories; (receipt …) comes last.

During the task, as soon as something durable happens:

At the end of the task:

consolidate(runId="issue-482", goal="Rework invoice numbering")

Once the outcome is known (CI finished, the user approved):

report_outcome(runId="issue-482", receiptId="<receipt from the search_memory output>",
               buildSuccess=true, testPassRate=0.97, userRating=5)

4. What happens over time?Copy Anchor Link

• Writing the same decisions again doesn't duplicate them; it strengthens their evidence.
• New corrections supersede old decisions; old information drops out of search but can still be traced with explain_memory.
• Patterns approved in successful tasks become procedures and appear in the procedures section for similar tasks.
• Failed tasks feed an avoid list; facts from low-scoring runs are flagged for review.
• If the same decision shows up in several projects, it becomes a tenant-wide rule candidate; if an admin approves it, it appears in all projects.

5. AuditingCopy Anchor Link

• See memories, supersession chains and history on the console's Memories page.
• To see why the agent recalled something, use search_memory(..., explain=true).
• To ask where a piece of information came from, use explain_memory(memoryId).

Responsibilities: product and applicationCopy Anchor Link

Related pagesCopy Anchor Link

• How memory works
• Consolidation, episodes and procedures
• Blocks and user profiles