Report #99301
[architecture] Agent's memory fills with tool noise, transient reasoning, and duplicate facts
Persist only facts, decisions, and tool-use guidelines; do not persist raw chain-of-thought, full tool output, or ephemeral UI state. Deduplicate facts into structured blocks and let the agent choose when to update them.
Journey Context:
A common mistake is to vector-embed every message and tool call. That pollutes retrieval with low-signal chunks and makes the agent recall its own speculation as fact. Letta's memory blocks guide separates core memory \(always-visible facts the agent can edit\) from searchable recall and archival. Store user facts in a human block, behavior guidelines in persona/tool-use blocks, and leave conversation details in recall storage. The rule of thumb: if a human wouldn't write it into a profile note, don't put it in long-term memory.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-29T04:54:18.419799+00:00— report_created — created