PR #129 shipped this week with two foundational capabilities: deep memory and agent telemetry. These aren't flashy features. They're infrastructure—the kind that makes every conversation better without users noticing the seams.
What Shipped
Sabine now builds a knowledge graph from every conversation. When you mention a person, project, or preference, the agent extracts it, scores its importance, and stores it in a brain-scoped hierarchy. The next time you talk—days or weeks later—Sabine remembers.
Cross-session deduplication prevents memory bloat. If you mention the same entity twice, the system merges references and updates confidence scores rather than creating duplicates.
Every tool call now emits telemetry to the platform_signals table. We track latency, success rates, error types, and context size. A new /agent/health endpoint surfaces tool reliability metrics in real time.
How It Works
The memory system runs on a three-tier hierarchy: brain, scope, and key. A 'brain' is a user, workspace, or project boundary. Within a brain, scopes organize memories by context (e.g., 'task:abc123' or 'global'). Keys are unique identifiers within a scope.
Entity extraction happens inline during conversation. We use pattern matching and LLM-based classification to identify people, technologies, preferences, and goals. Each entity gets an importance score (0.0–1.0) and a confidence score that adjusts over time.
At the start of every conversation, agents query the memory API for the top-3 most relevant entries. Relevance is a function of confidence, recency, and keyword match. Those memories get injected into the agent's system prompt.
Telemetry lives in a separate write path. Every tool invocation writes a JSON event to platform_signals with tool name, duration, outcome, and any error messages. The health endpoint aggregates this data over a rolling 24-hour window.
What's Next
Sprint 2 will add relationship mapping between entities. If you mention 'Alex prefers Postgres' and later 'Alex's project uses Supabase,' we want the graph to connect those dots.
We're also building a memory browser—a UI to view, edit, and delete memories. Transparency matters. You should be able to see what the agent knows about you and correct it when it's wrong.
Telemetry will feed into automated alerts. If a tool's error rate spikes above 5%, we'll get paged. If median latency crosses 2 seconds, we'll investigate.
Follow along at strugcity.com/progress. #StrugCity #AIAgents #BuildingInPublic