agentdb-recall

# Recall Pull relevant past memories into the current context. ## When to use - **Task start** — auto-load any prior episodes / patterns / skills that look relevant - **When stuck** — search for past failures + critiques on similar tasks - **User asks** — "have we done this?", "what do we know about X?" - **Decision point** — surface successful strategies before committing to one ## Pick the right tool | Need | Tool | Returns | |---|---|---| | Similar past tasks | `agentdb_reflexion_recall` | Episodes, top-k by similarity | | Lessons from past failures | `agentdb_critique_summary` | Combined critique text | | What worked last time | `agentdb_success_strategies` | Approach summaries from high-reward episodes | | Generic patterns / facts | `agentdb_pattern_search` | Patterns ranked by similarity | | Reusable skills by intent | `agentdb_skill_search` | Skills ranked by precondition match | ## Standard recall flow at task start 1. Embed the current task description. 2. Call `agentdb_reflexion_recall` with `k=5`, `minReward=0.5`. 3. Call `agentdb_pattern_search` with the same query, `k=5`. 4. Dedupe + rank — present the top 3-5 to the user as "things I found that look relevant." 5. If user-confirmed useful: call `agentdb_record_feedback` (or `recordFeedback` via the library) so the bandit learns. ## Filters worth knowing - `minReward` — drop low-quality matches (default 0.3). - `onlyFailures` — explicitly query the postmortem set when debugging. - `onlySuccesses` — only winning approaches when copying a strategy. - `timeWindowDays` — recent context only when the codebase has shifted. ## Don't - Don't dump all 50 hits — pick 3-5. Cognitive overload tanks quality. - Don't recall and ignore. If you used a memory, record feedback. If you didn't, record negative feedback. The bandit needs the signal. - Don't recall episode content verbatim into prompts when the task changes. Summarize the *lesson*, not the artifact.