@yemi33/minions 0.1.2401 → 0.1.2403

package/docs/named-agents.md

@@ -190,7 +190,10 @@ The capability matrix for each runtime is in [`docs/runtime-adapters.md`](runtim
 
 Each named agent has a personal memory file at
 `knowledge/agents/<agentId>.md` (e.g. `knowledge/agents/dallas.md`) — a personal notebook
-injected into that agent's prompt on every dispatch. The format convention is documented in
+whose entries are also indexed for task-aware SQL/FTS5 recall. In the legacy
+fallback it is injected into that agent's prompt; when active retrieval returns
+evidence, selected records replace the broad notebook appendix. The format
+convention is documented in
 [`knowledge/agents/README.md`](../knowledge/agents/README.md).
 
 **Written exclusively by the consolidation pipeline.** `appendToAgentMemory` in
@@ -199,16 +202,20 @@ to its author's file. Authorship comes from the inbox note's YAML frontmatter `a
 field, with the **filename prefix** (`<agent>-…md`) as fallback (`extractInboxAgent`). Key
 properties:
 
-- **Append-only** — agents must not edit their own memory files directly; write to
-  `notes/inbox/` and let the sweep route it.
-- **~25 KB cap** (`AGENT_MEMORY_BUDGET_BYTES`) — oldest sections pruned at section
-  boundaries to stay under budget.
+- **Pipeline-owned** — agents must not edit their own memory files directly; write
+  to `notes/inbox/` and let consolidation append, reconcile, prune, or summarize it
+  under the file lock.
+- **25,000-byte and entry-count caps** — canonical sections are pruned oldest-first
+  to the UTF-8 byte budget (`AGENT_MEMORY_BUDGET_BYTES`) and
+  `engine.agentMemoryMaxEntries` (default `300`).
 - **Only configured `config.agents` keys get a file** — the agent must be in the known-team
   set; `temp-*` ids are skipped. This is a strict superset of broadcast consolidation: the
   shared `notes.md` digest still happens; per-agent routing is *in addition*.
 
-**Prompt injection order.** [`engine/playbook.js`](../engine/playbook.js) injects, in this
-order, `pinned.md` → `notes.md` → `knowledge/agents/<agentId>.md` (when present). Missing
-files are silently skipped. The per-agent file is fenced as `<UNTRUSTED-INPUT>` and capped
-to the notes prompt budget before injection. The companion read-side reference is
+**Prompt memory modes.** [`engine/playbook.js`](../engine/playbook.js) always
+places non-empty `pinned.md` first. Default active retrieval then injects a
+bounded `Relevant Memory` pack when it finds evidence. If retrieval is disabled,
+shadowed, empty, or fails, the engine instead injects bounded `notes.md` followed
+by `knowledge/agents/<agentId>.md` when present. Every memory appendix is fenced
+as `<UNTRUSTED-INPUT>`. The full lifecycle and authority model are in
 [`docs/team-memory.md`](team-memory.md).