Workshop: Agent Memory Design
Question
What needs to change around Designing a Memory System for LLM-Based Agents now that the first design study exists as a durable synthesis note?
Why this workshop is back
The original agent-memory-design workshop closed on 2026-04-23 by promoting its coherent argument into kb/notes/designing-agent-memory-systems.md with the synthesis trait. That was the right closure for the first pass: the workshop had become one durable design study.
The note is now active again as an object of discussion. New questions are appearing around scope, connect-report application, silent-failure extraction, store-everything assumptions, and whether the architecture is software-specific or general across agent work surfaces. Those questions should not bloat the library note directly. They need a workshop where alternatives, objections, and candidate revisions can accumulate before promotion.
This reopened workshop is therefore a continuation workshop, not a rollback of the prior closure. The promoted note remains the current library artifact; this space holds discussion material around its next revision or extracted companion notes.
Current artifact
- Designing a Memory System for LLM-Based Agents — current synthesis note under discussion.
- Connection report — applied on 2026-04-23 as outbound links, reverse links, and index placements.
- Input-vs-output-driven memory — upstream workshop asking whether memory design should start from observed inputs or output requirements.
- Life-cycle management — records the first closure of
agent-memory-designas a single-file promotion with asynthesistrait.
Discussion threads
- Scope generality. The note has moved from software-project examples toward a broader "agent work surface" framing. Check whether that generality is earned or whether some mechanisms still depend on software-specific artifacts.
- Store-everything premise. Clarify how this design relates to output-driven memory. Is store-everything a capture substrate, a temporary bootstrap, or a stable architectural commitment?
- Observation taxonomy. Decide whether silent failures belong as a first-class extraction type alongside corrections, preferences, procedures, and discoveries, or whether they should sit under a broader operational-health category.
- Activation machinery. Make the typed cue index more operational: what fields, matching rules, and priority budgets are minimally required?
- Promotion oracles. Separate signals that identify candidates from signals that justify promotion into durable artifacts.
- Companion-note extraction. Identify claims inside the synthesis note that deserve standalone notes because they are being reused as premises elsewhere.
Working conventions
- Keep exploratory objections and alternatives here until they are stable enough to revise the note or extract a companion artifact.
- Do not link from library notes into this workshop; promote first.
- Prefer dated discussion files when a thread gets substantial, for example
2026-04-23-scope-generalization.md. - When the workshop produces a durable change, record it under "Graduated changes" below before closing or continuing.
Graduated changes
- None yet for the reopened workshop.
What would close this workshop
This continuation closes when either:
- the next revision of
designing-agent-memory-systems.mdabsorbs the active discussion threads and no residual questions remain, or - the active threads split into standalone notes/workshops and this directory no longer has a coherent center.