@leadbay/mcp 0.32.6 → 0.33.1

package/CHANGELOG.md

@@ -1,5 +1,67 @@
 # Changelog — @leadbay/mcp
 
+## 0.33.1 — 2026-09-02
+
+Follow-up to 0.32.0 (product#4007). A review finding landed after the merge.
+
+- **Only a REFUSAL counts as a commit failure.** The detached finisher's catch
+  recorded ANY rejection from the `pollPreprocess` → `commitMappings` chain as a
+  permanent commit failure — including the `ImportPhaseTimeout` raised when
+  preprocess still hasn't finished after the finisher's own 10-minute window.
+  That is not a backend refusal; the import is merely slow and may yet finish.
+  The registry has no expiry, so `leadbay_import_status` would have answered
+  `failed` for ever on a healthy import — worse than the forever-polling 0.32.0
+  set out to remove. A budget timeout now leaves the status at `committing`,
+  which is the honest reading; only an error from `commitMappings` itself is
+  recorded.
+- The finisher's window is deliberately not caller-controllable — a caller
+  shrinking it is the bug that budget exists to prevent — so the test reaches it
+  through a named module seam beside the existing `clearCommitFailures`, rather
+  than a parameter that would leak into the tool's public schema.
+
+## 0.33.0 — 2026-09-01
+
+Agent memory is removed (product#3996). The three `leadbay_agent_memory_*`
+tools, `packages/core/src/agent-memory/`, the `agent-memory://summary`
+resource, the `_meta.agent_memory` decoration and the `AGENT_MEMORY` server
+instruction are all deleted.
+
+**Why it goes rather than gets fixed.** The store was written when the MCP ran
+on the user's own machine and no host had memory of its own. Both premises are
+dead. Claude Cowork projects carry persistent memory into their scheduled
+tasks, Claude chat memory shipped to all plans, ChatGPT has memory plus remote
+connectors, and Codex reads the user's own file. Meanwhile the hosted server
+writes to a container layer that Argo replaces on every image build, so ours
+was strictly worse than the host's on the surface where it mattered most.
+
+**Measured before deciding.** PostHog `mcp tool called`, tools matching
+`%memory%`, 120 days, whole fleet: **121 captures, 30 recalls, 13 users.**
+79 of the 121 captures came from one hosted IP and stop on 2026-07-29. August
+was 11 captures and 3 recalls. The feature wrote four times more than it was
+read and almost nothing read it twice.
+
+**What replaces it.** Reading that user's verbatim `triggered_by` sentences
+rather than a summary of them, the largest group by far was targeting — *"je
+vise en priorité les entreprises qui ont plus de 100 voitures dans leur parc"*,
+*"les transporteurs ne sont pas forcément de bons prospects sauf ceux qui font
+du last mile delivery"*. Stored as notes those changed one conversation. The
+new `headers/durable-preferences` snippet, included by the seven prompts that
+carried the memory preamble, routes them to `leadbay_refine_prompt` instead,
+where they mutate `computing_intelligence` and change what the product finds
+for the whole organization. A rejected lead goes to a dislike.
+
+- `withAgentMemoryMeta` is unwrapped from nine composites. It called
+  `resolveMe()` on every invocation, so each of those tools loses a round-trip.
+- `LEADBAY_AGENT_MEMORY`, `EV_AGENT_MEMORY_*` and the three
+  `captureAgentMemory*` telemetry methods are gone. Historical events stay in
+  PostHog; nothing new is emitted.
+- `assembler.ts` no longer injects a memory pointer into routed tool
+  descriptions, and `memory_protocol` is removed from the frontmatter schema.
+- Deliberately NOT landed: `agent-memory/durability.ts`, a boot-time warning
+  for an ephemeral memory root written while this was still going to be fixed
+  with a volume. With nothing writing to disk the condition it detects can no
+  longer occur, so it is not a safety net, it is dead code.
+
 ## 0.32.6 — 2026-09-01
 
 The 401 auto-retry is GET-only — replaying a write could double-execute a

package/README.md

@@ -22,26 +22,18 @@ A Model Context Protocol server that lets Claude Desktop, Cursor, Claude Code, a
 
 > **Upgrading?** See [CHANGELOG](../../CHANGELOG.md) and [MIGRATION.md](./MIGRATION.md) for version-specific behavior changes. Key callout: composite write tools are **ON by default** since 0.3.0 — set `LEADBAY_MCP_WRITE=0` to restore read-only behavior.
 
-## Agent memory
-
-Leadbay MCP keeps a local, per-account agent memory at
-`~/.leadbay/memory/{account_id}/`. It stores append-only JSONL learnings
-about user taste signals such as preferred sectors, regions, deal size,
-communication style, and qualification rules.
-
-The memory tools are always exposed:
-
-- `leadbay_agent_memory_recall` reads the consolidated top signals.
-- `leadbay_agent_memory_capture` appends a new learning after the user reveals
-  a material preference.
-- `leadbay_agent_memory_review` lists entries and gates retractions or org
-  promotion through user confirmation.
-
-The main leads-touching tools (`leadbay_account_status`,
-`leadbay_pull_leads`, `leadbay_pull_followups`,
-`leadbay_prepare_outreach`, `leadbay_research_lead_by_id`) also attach
-`_meta.agent_memory.summary` automatically. Set `LEADBAY_AGENT_MEMORY=off`
-to suppress this ambient metadata.
+## What Leadbay remembers
+
+Leadbay stores nothing about how you like to work. Your assistant already does
+that — Claude, ChatGPT and Codex each keep their own memory of your tone,
+naming and habits, and it survives across conversations without us.
+
+What Leadbay stores is what changes **who it finds for you**. When you tell the
+agent something about your market ("I target fleets over 100 vehicles",
+"carriers are a bad fit unless they do last-mile"), it calls
+`leadbay_refine_prompt`, and that reshapes your recommendations for your whole
+organization — in the web app too, on every refresh. A lead you reject is
+recorded as a dislike, which feeds the recommender the same way.
 
 ## 1. Install