Agent Memory & Codebase Knowledge Base

Agent Memory is a persistent, self-improving knowledge layer built into the autonomous development pipeline. Rather than starting from scratch on every pipeline run, implementation agents accumulate and reuse project-specific knowledge — conventions, past mistakes, gotchas, and successful patterns — so code quality compounds over time.

Overview

Every pipeline run now:

1. Loads relevant memory entries ranked by confidence and helpfulness
2. Injects them into the implementation agent's system prompt
3. Records which entries were used

Every night at 3 AM UTC, the Distiller scans the last 24 hours of pipeline logs, extracts new patterns, and reinforces entries that repeat — automatically raising their confidence score.

---

Entry Structure

Each memory entry has the following fields:

Field	Description
type	One of: pattern, mistake, convention, dependency, gotcha, success
category	One of: schema, api, ui, testing, auth, performance, inngest, deployment, general
title	Short descriptive title
content	Full explanation
codeExample	Optional illustrative code snippet
affectedFiles	Optional list of relevant file paths
tags	Optional free-form tags
confidence	Float 0–1. Higher = more likely to be injected
humanVerified	Boolean. Verified entries get confidence 0.95
timesUsed	How many times this entry has been surfaced to agents
source	agent, manual, or distiller

Entry Types

Type	Icon	Purpose
pattern	⚡	Recurring coding patterns worth following
mistake	⚠️	Known errors to avoid
convention	📖	Project-specific code conventions
dependency	📦	Dependency usage notes
gotcha	🚨	Tricky edge cases that have caused issues
success	✅	Approaches that have worked well

---

Confidence Scoring

Confidence controls whether an entry gets injected into agent prompts:

• 0.0–0.49 — low confidence; entry exists but is rarely injected
• 0.50–0.79 — moderate confidence; injected when relevant
• 0.80–1.0 — high confidence; consistently injected

Reinforcement learning: when the distiller observes the same pattern again, it calls upsertMemoryEntry() which increments the confidence score. Repeated mistakes or confirmed patterns gradually earn higher confidence without manual intervention.

Human verification: clicking Verify in the dashboard immediately sets confidence to 0.95.

---

The Nightly Distiller

The Inngest cron (agentMemoryDistiller) runs every night at 3:00 AM UTC. It:

1. Queries completed and failed pipeline runs from the last 24 hours
2. Parses agent log output for known error patterns
3. Creates or reinforces structured memory entries

Automatically Detected Patterns

Log Signal	Memory Entry Created
TypeScript TS2339 error	mistake — property does not exist on type
DynamicServerError	gotcha — page must use force-dynamic
Drizzle .where() chaining	mistake — incorrect query builder usage
Inngest Jsonify error	gotcha — return type serialisation issue
Missing module import error	mistake — import path or package missing
New router/schema/Inngest file written	convention — established file structure pattern
PR merged successfully	success — approach that led to green CI

---

Dashboard

Navigate to Products → [Product] → Agent Memory (/dashboard/products/[id]/agent-memory) to view and manage the knowledge base.

Stats Cards

• Total Entries — all active memory entries
• Patterns — entries of type pattern
• Mistakes Caught — entries of type mistake
• Human Verified — entries with humanVerified: true

Filtering

Filter the entry list by type (pattern, mistake, convention, etc.) and category (schema, api, ui, etc.) using the dropdowns at the top of the list.

Entry Cards

Each card shows:

• Title, type badge, category badge, confidence percentage
• A Verified badge if humanVerified is true
• Usage count and source
• Expandable More section with code example, affected files, and tags

Actions

• Verify (shield icon) — sets humanVerified: true and confidence to 0.95
• Delete (trash icon) — soft-deletes the entry (active = false); it will no longer be injected

Adding Entries Manually

Click Add Entry to open the creation dialog. Manual entries are created with confidence: 0.9 and humanVerified: true by default, so they are injected immediately.

Required fields:

• Type — select from the six types
• Category — select from the nine categories
• Title — short label
• Content — full explanation

Optional fields:

• Code Example — a snippet illustrating the point
• Tags — comma-separated

---

tRPC API Reference

All endpoints are under the agentMemory router.

agentMemory.list

Returns paginated memory entries for a project.

Input

{
  projectId: string;
  type?: "pattern" | "mistake" | "convention" | "dependency" | "gotcha" | "success";
  category?: "schema" | "api" | "ui" | "testing" | "auth" | "performance" | "inngest" | "deployment" | "general";
  cursor?: number; // offset
  limit?: number;
}

agentMemory.stats

Returns aggregate counts for a project's memory entries.

Input

{ projectId: string }

agentMemory.create

Creates a human-authored entry (confidence: 0.9, humanVerified: true).

Input

{
  projectId: string;
  type: EntryType;
  category: EntryCategory;
  title: string;
  content: string;
  codeExample?: string;
  affectedFiles?: string[];
  tags?: string[];
}

agentMemory.update

Updates any field on an existing entry.

Input

{ id: string; [field]: value }

agentMemory.delete

Soft-deletes an entry (sets active = false).

Input

{ id: string }

agentMemory.verify

Marks an agent-extracted entry as human-verified and sets confidence to 0.95.

Input

{ id: string }

---

Library Functions

These are used internally by the pipeline and distiller. You do not need to call them directly unless building custom tooling.

getRelevantMemory(projectId, options?)

Retrieves the top-N active entries ranked by confidence and usage, optionally filtered by category.

formatMemoryForInjection(entries)

Renders an array of entries into a compact markdown block grouped by category, suitable for inclusion in an LLM system prompt. Prefixes each entry with a type icon.

upsertMemoryEntry(projectId, entry)

Inserts a new entry or, if a matching title already exists, increments its confidence score (reinforcement learning).

recordMemoryUsage(ids)

Increments timesUsed for the given entry IDs.

---

Pipeline Integration

Memory loading is best-effort and non-blocking — if the memory query fails, the pipeline continues unaffected.

Orchestrator (pipeline.ts)
  └─ getRelevantMemory(projectId)
       └─ formatMemoryForInjection(entries)
            └─ passed as agentMemory in baseEventData
                 └─ implementation agent reads event.data.agentMemory
                      └─ injected into system prompt via agentMemory option

The PipelineEventData type now includes:

interface PipelineEventData {
  // ... existing fields
  agentMemory?: string; // formatted markdown block
}