Agents shouldn't know they have memory: background 'dreaming' beats save_memory tools
uriwa · reddit · 2026-11-22
Reddit user uriwa argues that giving agents explicit savememory tools or letting them rewrite their own system prompt fails fast in practice: models miss subtle preferences, hoard ephemeral junk, and narrate their memory mechanics to users.
The working alternative treats long-term memory like sleep rather than note-taking:
- Consolidation is tied to history compaction—a lightweight background pass compares new summaries against the base prompt and a separate learned memory block.
- Only durable items merge (standing preferences, recurring workflows, key contacts, superseding corrections); stale noise is pruned and the base prompt is never touched.
- The agent is never told the mechanism exists—learned memory just appears in context.
- Hard scoping: only shared-context personal/team agents, consolidated only from verified 1:1 owner conversations, so external contacts can't prompt-inject permanent rules into memory.
- The learned memory shows as a separate editable section in the dashboard.
More from coding & agent
- Dev shows ASC CLI maintained daily by Opus 5.5 in 'For Agents, By Agents' talk — rudrank · 2026-11-22
- Gemini subscriptions may work in third-party agent harnesses starting December 2 — testingcatalog · 2026-11-22
- Agno's AgentOS guide: build a team agent platform from zero to production — pritisinghhhh · 2026-11-22
- Background memory consolidation tied to compaction, with 3.5k-char budget and prompt-injection safeguards — uriwa · 2026-11-22
- Proactive Grok Bot setup: 28 of 34 weekly stops produce zero notifications — marcelmarais · 2026-11-22
- Anthropic engineer: 90% of staff built self-prompting agent systems around Opus — joemeno · 2026-11-22