@lebronj/pi-suite 0.1.16 → 0.1.18

package/skills/pi-skill/SKILL.md

@@ -27,7 +27,9 @@ When adding, removing, renaming, or materially changing any pi capability in thi
 
 Installed memory capability comes from `@jhp/pi-memory`, vendored inside `@lebronj/pi-suite` and also usable as a standalone pi package.
 
-Memory files live under `~/.pi/agent/memory/`:
+Memory files live under the resolved memory root. Standalone fallback is `~/.pi/agent/memory`; Multica-connected runs can resolve to `~/multica_workspaces/<workspace_id>/.pi/agents/<agent_id>/memory` with skill drafts under that agent's `skills/drafts`.
+
+Memory files include:
 
 - `MEMORY.md`: durable facts, decisions, and preferences.
 - `USER.md`: structured user profile and stable preferences.
@@ -38,7 +40,8 @@ Memory files live under `~/.pi/agent/memory/`:
 - `.curator-state.json`: last curator run state.
 - `.curator-service.json`: external curator service state.
 - `audit/curator.jsonl`: curator audit trail.
-- `~/.pi/agent/skill-drafts/<slug>/SKILL.md`: disabled skill drafts created after explicit approval.
+- Resolved skill draft root `<slug>/SKILL.md`: disabled skill drafts created after explicit approval.
+- Multica agent roots also contain `inbox/`, `shared-cache/`, `skills/generated/`, `profile/`, `feedback/feedback.jsonl`, and `sync_queue/`.
 
 Memory tools:
 
@@ -47,10 +50,15 @@ Memory tools:
 - `memory_edit`: read/add/replace/remove/replace_all/compact structured entries in `MEMORY.md`, `USER.md`, `STATE.md`, and `REVIEW.md`.
 - `scratchpad`: add/done/undo/clear/list checklist items.
 - `memory_search`: qmd-backed keyword, semantic, or deep search across memory files.
-- `memory_curate`: manually run curator lifecycle rules and scan yesterday's daily log into `REVIEW.md` when learning is enabled.
+- `memory_curate`: manually run curator lifecycle rules and scan yesterday's daily log into `REVIEW.md` when learning is enabled; output includes pending proposal counts.
 - `memory_learning_approve`: approve a proposed memory promotion or disabled skill draft by exact id.
 - `memory_learning_reject`: reject or archive a review candidate/proposal without deleting it.
 - `memory_skill_drafts`: list proposed skill drafts.
+- `memory_skill_list`: list current-agent draft, generated, and enabled memory-managed skills.
+- `memory_skill_enable`: explicitly enable a `draft:<slug>` or `generated:<id>` skill by copying it into `skills/enabled/<skill-name>/` and auditing the action.
+- `memory_skill_disable`: remove an enabled skill copy while preserving its draft/generated source.
+- `/memory-skill`: slash command to list/enable/disable current-agent memory-managed skills.
+- `/memory-review`: slash command to list/show/approve/reject/archive pending memory and skill proposals in the current resolved root.
 - `memory_curator_enable`: enable the external daily curator service using systemd user timer or cron fallback.
 - `memory_curator_disable`: disable and uninstall the external daily curator service.
 - `memory_curator_status`: show service backend, schedule, and state.
@@ -59,6 +67,11 @@ Memory tools:
 - `memory_version_list`: list recent snapshots.
 - `memory_version_restore`: restore `memory`, `skill-drafts`, or `all` from a snapshot id after creating a pre-restore backup.
 - `memory_version_push`: manually push the local evolution repo to GitHub.
+- `memory_sync_upload` / `/memory-sync-upload`: upload governed candidates, profiles, and feedback to Multica when remote config is set.
+- `memory_sync_pull` / `/memory-sync-pull`: pull only current-agent deliveries into local inbox/cache/generated-skill directories.
+- `memory_feedback`: append injected/used/ignored/success/failure/conflict feedback for a shared unit.
+- `memory_curator_manager_mark_dirty` and `memory_curator_manager_scan` / `/memory-curator-manager-scan`: register dirty roots and process them through the singleton Local Curator Manager.
+- `memory_curator_manager_enable`, `memory_curator_manager_status`, `memory_curator_manager_disable` and `/memory-curator-manager-*` commands: manage the singleton manager service; default schedule is every 6 hours and it only processes dirty roots.
 
 Structured metadata example:
 
@@ -77,12 +90,17 @@ Curator and learning behavior:
 - Session shutdown may extract conservative learning candidates into `REVIEW.md`; they are not injected as normal memory and are not auto-enabled.
 - `memory_curate` scans yesterday's daily log once per content hash into review candidates, then curator lifecycle and proposal rules process those candidates.
 - Repeated candidates can become proposed memory promotions or proposed disabled skill drafts after `memory_curate`.
-- Approval is explicit by default: memory proposals write to memory stores; skill proposals write disabled drafts under `~/.pi/agent/skill-drafts/`.
+- Approval is explicit by default: memory proposals write to memory stores; skill proposals write disabled drafts under the resolved skill draft root.
+- Draft and generated skills stay disabled until `memory_skill_enable` copies them into `skills/enabled`; enabled skills are injected as `<available_skills>` metadata for the current agent.
+- Pi session start can show one pending-review hint; disable with `PI_MEMORY_REVIEW_STARTUP_HINT=0`.
+- Local multi-agent self-evolution supports one Local Curator Manager registry/dirty-root API for many agent roots, plus a manager service that runs `manager-scan` every 6 hours and exits quickly when no root is dirty.
+- The local loop also covers share candidate queue, profile generation, sync upload/pull, downflow receive cache, generated skills, enabled skill lifecycle, and feedback JSONL helpers.
+- Server downflow is per-Agent delivery, not broadcast; local receive writes only `inbox/`, `shared-cache/`, or `skills/generated/` and never overwrites formal memory or auto-enables skills.
 - The curator avoids semantic auto-delete/merge; ambiguous learning stays in review first.
 
 Memory versioning:
 
-- Runtime memory remains authoritative at `~/.pi/agent/memory`; disabled skill drafts remain authoritative at `~/.pi/agent/skill-drafts`.
+- Runtime memory remains authoritative at the resolved memory root; disabled skill drafts remain authoritative at the resolved skill draft root.
 - Versioning mirror and snapshots live at `~/.pi/agent/evolution` by default.
 - No remote is configured by default; memory evolution stays local per user/machine unless the user adds a personal private remote.
 - Automatic local snapshot + commit is enabled by default; automatic push is disabled unless `PI_EVOLUTION_AUTO_PUSH=1`.
@@ -114,6 +132,10 @@ QMD search:
 Useful memory environment variables:
 
 - `PI_MEMORY_DIR`: override memory storage directory.
+- `PI_SKILL_DRAFTS_DIR`: override disabled skill draft root.
+- `PI_AGENT_ROOT`: current local agent root; derives memory, skill drafts, inbox, shared cache, profile, feedback, and sync queue.
+- `MULTICA_WORKSPACE_ID`, `MULTICA_AGENT_ID`, `MULTICA_WORKSPACES_ROOT`, `MULTICA_RUN_ID`: Multica scoped root and feedback context.
+- `PI_AGENT_INBOX_DIR`, `PI_AGENT_SHARED_CACHE_DIR`, `PI_AGENT_PROFILE_DIR`, `PI_AGENT_FEEDBACK_DIR`, `PI_AGENT_SYNC_QUEUE_DIR`: override agent subdirectories.
 - `PI_MEMORY_SNAPSHOT`: `stable` or `per-turn` context injection mode.
 - `PI_MEMORY_QMD_UPDATE`: `background`, `manual`, or `off`.
 - `PI_MEMORY_NO_SEARCH=1`: disable per-turn search injection.
@@ -123,6 +145,8 @@ Useful memory environment variables:
 - `PI_MEMORY_AUTO_APPROVE_MEMORY=1`: automatically approve newly created memory proposals.
 - `PI_MEMORY_AUTO_APPROVE_SKILL_DRAFTS=1`: automatically create newly proposed disabled skill drafts.
 - `PI_MEMORY_CURATOR_STARTUP_HINT=0`: hide the disabled-curator startup hint.
+- `PI_MEMORY_REVIEW_STARTUP_HINT=0`: hide pending review proposal startup hints.
+- `PI_MEMORY_REMOTE_URL` and `PI_MEMORY_REMOTE_TOKEN`: enable Multica candidate/profile/feedback upload and current-agent delivery pull.
 - `PI_EVOLUTION_ENABLED=0`: disable snapshot + git versioning.
 - `PI_EVOLUTION_DIR`: override evolution repo directory; default `~/.pi/agent/evolution`.
 - `PI_EVOLUTION_REMOTE`: optional personal private Git remote; unset by default.
@@ -190,9 +214,10 @@ Autogoal is provided by `autogoal.ts`.
 Goal mode is provided by `goal-mode.ts`.
 
 - Start with `/goal <objective>`.
-- It injects hidden goal context, enables the `goal` tool, and auto-continues until the objective is complete, paused, dropped, blocked, or interrupted.
+- It injects hidden goal context, enables the `goal` tool, and auto-continues until the objective is complete, paused, dropped, budget-limited, or interrupted.
+- It tracks assistant token usage and elapsed time, supports `/goal budget <tokens|off>`, and treats budget exhaustion as not completion.
 - The agent must verify current files/checks before calling `goal({ op: "complete" })`.
-- Useful commands: `/goal show`, `/goal pause`, `/goal resume`, `/goal drop`, `/goal auto on`, `/goal auto off`.
+- Useful commands: `/goal show`, `/goal pause`, `/goal resume`, `/goal drop`, `/goal budget <tokens|off>`, `/goal auto on`, `/goal auto off`.
 
 ## Pet Companion
 

package/vendor/pi-memory/README.md

@@ -4,15 +4,15 @@ Structured, time-aware memory extension for pi. It stores memory as plain Markdo
 
 ## Features
 
-- Plain Markdown storage under `~/.pi/agent/memory`.
+- Plain Markdown storage under resolved memory roots: standalone `~/.pi/agent/memory`, explicit `PI_MEMORY_DIR`, or Multica `~/multica_workspaces/<workspace>/.pi/agents/<agent>/memory`.
 - Unified tools for memory write/read/edit/search and scratchpad management.
 - Structured `USER.md`, `STATE.md`, and `REVIEW.md` entries with metadata.
 - Optional qmd-powered keyword, semantic, and deep search.
 - KV cache-stable context injection by default.
 - Curator core for exact dedupe, event lifecycle updates, temporary review, quota reset, and audit logs.
-- Review-first learning candidates that can become memory promotions or disabled skill drafts after approval.
+- Review-first learning candidates that can become memory promotions or disabled skill drafts after approval, with `/memory-review` pending reminders.
+- Local multi-agent self-evolution layout for Multica: scoped memory/skill roots, `sync_queue/`, `inbox/`, `shared-cache/`, `profile/`, and `feedback/`.
 - Optional external curator service via systemd user timer or cron so curation can run even when pi is closed.
-- Snapshot + git versioning for `memory/` and disabled `skill-drafts/`, with restore support.
 
 ## Installation
 
@@ -39,22 +39,26 @@ The extension auto-creates the `pi-memory` qmd collection and path contexts on s
 ## File Layout
 
 ```text
+# Standalone fallback
 ~/.pi/agent/memory/
-  MEMORY.md              # Durable facts, decisions, and preferences
-  USER.md                # Structured user profile and stable preferences
-  STATE.md               # Current dated state, events, temporary facts, quotas
-  REVIEW.md              # Review queue for stale or merge-candidate memories
-  SCRATCHPAD.md          # Checklist of open items
-  .curator-state.json    # Last curator run state
-  audit/curator.jsonl    # Curator audit trail
-  daily/YYYY-MM-DD.md    # Daily append-only logs
 ~/.pi/agent/skill-drafts/
-  <slug>/SKILL.md        # Disabled skill drafts created after approval
-~/.pi/agent/evolution/
-  memory/                # Current memory mirror
-  skill-drafts/          # Current skill draft mirror
-  snapshots/<id>/        # Point-in-time backup with manifest.json
-  manifests/<id>.json    # Snapshot manifest index
+
+# Multica-connected scoped root
+~/multica_workspaces/<workspace_id>/.pi/agents/<agent_id>/
+  memory/
+    MEMORY.md USER.md STATE.md REVIEW.md SCRATCHPAD.md
+    .curator-state.json
+    audit/curator.jsonl
+    daily/YYYY-MM-DD.md
+  skills/
+    drafts/<slug>/SKILL.md      # Disabled skill drafts created after approval
+    generated/                  # Downflow skill deliveries, not auto-enabled
+    enabled/                    # Reserved for explicit enablement
+  inbox/memory/ inbox/skills/   # Raw downflow deliveries for this agent only
+  shared-cache/memory/ shared-cache/skills/
+  profile/user-profile.md agent-profile.md task-profile.md capability-profile.md
+  feedback/feedback.jsonl
+  sync_queue/memory-candidates.jsonl skill-candidates.jsonl
 ```
 
 Structured entries are separated by `§` and may start with metadata:
@@ -88,14 +92,22 @@ Metadata keys currently supported by tools and curator rules:
 | `memory_learning_approve` | Approve one proposed memory or skill promotion by exact id |
 | `memory_learning_reject` | Reject or archive one review item by exact id |
 | `memory_skill_drafts` | List proposed skill drafts |
+| `memory_skill_list` | List current-agent draft, generated, and enabled skills |
+| `memory_skill_enable` | Enable a draft/generated skill into `skills/enabled` |
+| `memory_skill_disable` | Disable an enabled skill while preserving its source |
 | `memory_curator_enable` | Enable the external daily curator service |
 | `memory_curator_disable` | Disable the external daily curator service |
 | `memory_curator_status` | Show service backend, schedule, and state |
-| `memory_version_status` | Show local evolution git repo status |
-| `memory_version_snapshot` | Manually snapshot memory and skill drafts |
-| `memory_version_list` | List recent snapshots |
-| `memory_version_restore` | Restore `memory`, `skill-drafts`, or `all` from a snapshot id |
-| `memory_version_push` | Manually push the evolution repo to GitHub |
+| `memory_sync_upload` / `/memory-sync-upload` | Upload governed candidates, profiles, and feedback when Multica remote config is set |
+| `memory_sync_pull` / `/memory-sync-pull` | Pull current-agent deliveries into inbox/cache/generated-skill locations |
+| `memory_feedback` | Append shared unit usage feedback to `feedback/feedback.jsonl` |
+| `memory_curator_manager_mark_dirty` | Register and mark the current agent root dirty |
+| `memory_curator_manager_scan` / `/memory-curator-manager-scan` | Process dirty roots from the singleton manager registry |
+| `memory_curator_manager_enable` | Enable the singleton manager service; default checks dirty roots every 6 hours |
+| `memory_curator_manager_disable` | Disable the singleton manager service |
+| `memory_curator_manager_status` | Show singleton manager service status |
+| `/memory-skill` | Slash command for list/enable/disable skill lifecycle actions |
+| `/memory-review` | List/show/approve/reject/archive/compact pending memory and skill proposals in the current root |
 
 ### memory_write Targets
 
@@ -127,6 +139,8 @@ By default, pi-memory injects a stable snapshot into the system prompt. It inclu
 5. current `STATE.md` entries, excluding `status:past` and `status:archived`
 6. `MEMORY.md`
 7. Yesterday's daily log
+8. Matched shared-cache memories/generated skills for Multica-scoped roots
+9. Explicitly enabled current-agent skills as `<available_skills>` metadata
 
 The stable snapshot refreshes on session start, compaction, long-term or structured writes, and day rollover. Use `memory_read` or `memory_search` for the latest authoritative state.
 
@@ -170,45 +184,58 @@ When candidates repeat, pi-memory updates the existing candidate instead of appe
 
 Approval is explicit by default:
 
+- `memory_curate` reports pending memory/skill proposal counts and suggests `/memory-review` plus approve/reject tools.
+- Pi session start shows one lightweight pending-review hint when proposals exist; disable with `PI_MEMORY_REVIEW_STARTUP_HINT=0`.
+- `/memory-review` lists pending proposals and supports `show <id>`, `approve <id>`, `reject <id>`, and `archive <id>` for the current resolved root.
 - `memory_learning_approve` on a memory proposal writes `MEMORY.md`, `USER.md`, or `STATE.md` depending on the proposal target.
-- `memory_learning_approve` on a skill proposal writes `~/.pi/agent/skill-drafts/<slug>/SKILL.md` and marks the proposal approved.
+- `memory_learning_approve` on a skill proposal writes the current resolved skill draft root and marks the proposal approved.
 - Skill drafts are disabled. They are not moved into enabled skill directories automatically.
+- `memory_skill_enable` explicitly copies a `draft:<slug>` or `generated:<id>` skill into `skills/enabled/<skill-name>/` and writes `memory/audit/skill-lifecycle.jsonl`.
+- `memory_skill_disable` removes only the enabled copy; the draft/generated source remains for later review.
+- Enabled skills are injected as available-skill metadata so the agent can read the corresponding `SKILL.md` when the task matches.
 - `memory_learning_reject` marks a candidate or proposal as `rejected` or `archived` without deleting it.
 
 Old candidates are lifecycle-managed without deletion. Low-confidence candidates can become `archived`; others become `needs_review` first. `REVIEW.md` remains the evidence and audit trail, so approved items are marked rather than removed.
 
 Current learning extraction is text-based: it reads user/assistant conversation messages and asks the active model for structured candidates. It does not yet inspect structured tool-call graphs directly. Curator patch audit remains in `audit/curator.jsonl`; learning approvals are tracked through `REVIEW.md` proposal metadata and status changes.
 
-## Memory Versioning
+## Local Multi-Agent Self-Evolution
 
-Pi-memory mirrors the authoritative runtime directories into a local evolution repo and stores point-in-time snapshots before important changes:
+Resolver priority:
 
-```text
-~/.pi/agent/evolution/
-  memory/
-  skill-drafts/
-  snapshots/<snapshot-id>/
-    memory/
-    skill-drafts/
-    manifest.json
-  manifests/<snapshot-id>.json
-```
+1. `PI_MEMORY_DIR` and `PI_SKILL_DRAFTS_DIR` when explicitly set.
+2. `PI_AGENT_ROOT` when set, deriving `memory/` and `skills/drafts/`.
+3. `MULTICA_WORKSPACE_ID` + `MULTICA_AGENT_ID`, deriving `~/multica_workspaces/<workspace_id>/.pi/agents/<agent_id>/` (or `MULTICA_WORKSPACES_ROOT`).
+4. Standalone fallback `~/.pi/agent/memory` and `~/.pi/agent/skill-drafts`.
 
-Authoritative runtime data remains `~/.pi/agent/memory` and `~/.pi/agent/skill-drafts`; `~/.pi/agent/evolution` is a versioned mirror and backup repo.
+`MULTICA_MEMBER_ID` is reserved and does not change v1 paths. Agent A and Agent B therefore get separate memory, skill drafts, generated skills, profiles, feedback, and sync queues.
 
-Automatic hooks snapshot before and sync/commit after `memory_write`, mutating `memory_edit`, mutating `scratchpad`, `memory_curate`, learning approve/reject, session summary/handoff writes, compaction handoffs, and external `jhp-pi-memory-curator run-once`. `memory_curate` also scans yesterday's daily log into `REVIEW.md` when learning is enabled and that daily file changed since the last scan. Read-only operations do not snapshot. Snapshots use a sliding window and keep the latest 100 snapshots by default; creating snapshot 101 deletes the oldest snapshot directory and manifest.
+The package includes local primitives for the full local loop:
 
-Tools and slash commands:
+- `ensureAgentRoot()` initializes the scoped directory tree.
+- `markCurrentRootDirty()` and `scanDirtyRoots()` implement a single Local Curator Manager registry, manager-level locking, stale lock cleanup, and per-root `.curator.lock` processing.
+- `generateShareCandidatesFromReview()` and `appendEvolutionCandidate()` write governed share candidates to `sync_queue/` and block secret-like payloads.
+- `generateProfiles()` writes conservative local profiles for remote matching input.
+- `syncUpload()` / `memory_sync_upload` POST candidates, profiles, and feedback, using a checkpoint to avoid re-uploading prior candidate ids or feedback lines.
+- `syncPull()` / `memory_sync_pull` pull only current-agent deliveries and call `receiveDelivery()`.
+- `receiveDelivery()` writes server downflow only to `inbox/`, `shared-cache/`, or `skills/generated/`; it never overwrites formal memory or enables skills.
+- `appendFeedbackEvent()` / `memory_feedback` writes injected/used/ignored/success/failure/conflict events to `feedback/feedback.jsonl` for connector upload.
 
-- `memory_version_status` / `/memory-version-status`
-- `memory_version_snapshot` / `/memory-version-snapshot [reason]`
-- `memory_version_list` / `/memory-version-list`
-- `memory_version_restore` / `/memory-version-restore <snapshot-id> [memory|skill-drafts|all]`
-- `memory_version_push` / `/memory-version-push`
+Server delivery is per-agent matching, not broadcast. The local runtime must only pull deliveries for the current `MULTICA_AGENT_ID` and still filter before injection.
 
-Restore always creates a pre-restore snapshot first, then restores the selected target, syncs the mirror, and commits `memory: restore snapshot <id>`.
+### Local Curator Manager Service
+
+The manager service is separate from the standalone daily memory curator. It runs `jhp-pi-memory-curator manager-scan` against the registry and only processes roots marked `dirty`; when there are no dirty roots it exits after a cheap registry check.
+
+Default cadence is every six hours using cron syntax `0 */6 * * *`. Enable, inspect, or disable it with:
+
+```bash
+jhp-pi-memory-curator manager-enable
+jhp-pi-memory-curator manager-status
+jhp-pi-memory-curator manager-disable
+```
 
-No remote is configured by default. This keeps each user's memory evolution local to their machine and avoids writing personal memory into a shared team repo. If remote backup is desired, set `PI_EVOLUTION_REMOTE` to a personal private repo before setup, or add `origin` manually later. Auto commit is on; auto push is off unless `PI_EVOLUTION_AUTO_PUSH=1` is set. Any remote should be private because memory contents are committed in plaintext, including any secret accidentally written to memory.
+Pi tools provide the same flow: `memory_curator_manager_enable`, `memory_curator_manager_status`, and `memory_curator_manager_disable`. Slash commands `/memory-curator-manager-enable`, `/memory-curator-manager-status`, and `/memory-curator-manager-disable` are also available. Pass `--registry` or the tool `registry` parameter to override the default registry path.
 
 ## External Curator Service
 
@@ -220,8 +247,6 @@ CLI:
 jhp-pi-memory-curator enable --schedule 03:00
 jhp-pi-memory-curator status
 jhp-pi-memory-curator run-once
-jhp-pi-memory-curator snapshot --reason "manual backup"
-jhp-pi-memory-curator push
 jhp-pi-memory-curator disable
 ```
 
@@ -237,23 +262,29 @@ The controller uses a systemd user timer when available and falls back to cron.
 
 | Variable | Values | Default | Description |
 | --- | --- | --- | --- |
-| `PI_MEMORY_DIR` | path | `~/.pi/agent/memory` | Override storage directory |
+| `PI_MEMORY_DIR` | path | resolved fallback | Override memory root |
+| `PI_SKILL_DRAFTS_DIR` | path | resolved fallback | Override disabled skill draft root |
+| `PI_AGENT_ROOT` | path | derived from Multica env when present | Current local agent root |
+| `MULTICA_WORKSPACES_ROOT` | path | `~/multica_workspaces` | Base for derived Multica roots |
+| `MULTICA_WORKSPACE_ID` | string | unset | Current Multica workspace id |
+| `MULTICA_AGENT_ID` | string | unset | Current Multica agent id |
+| `MULTICA_RUN_ID` | string | unset | Current Multica run id for feedback |
+| `PI_AGENT_INBOX_DIR` | path | `$PI_AGENT_ROOT/inbox` | Downflow inbox override |
+| `PI_AGENT_SHARED_CACHE_DIR` | path | `$PI_AGENT_ROOT/shared-cache` | Shared cache override |
+| `PI_AGENT_PROFILE_DIR` | path | `$PI_AGENT_ROOT/profile` | Profile directory override |
+| `PI_AGENT_FEEDBACK_DIR` | path | `$PI_AGENT_ROOT/feedback` | Feedback directory override |
+| `PI_AGENT_SYNC_QUEUE_DIR` | path | `$PI_AGENT_ROOT/sync_queue` | Upload queue override |
+| `PI_MEMORY_MANAGER_SCHEDULE` | cron/systemd schedule | `0 */6 * * *` | Suggested schedule for the Local Curator Manager dirty-root scan service |
+| `PI_MEMORY_REVIEW_STARTUP_HINT` | `0`/unset | unset | Set `0` to hide startup pending review hints |
 | `PI_MEMORY_SNAPSHOT` | `stable`, `per-turn` | `stable` | Stable context injection or legacy per-turn rebuild |
 | `PI_MEMORY_QMD_UPDATE` | `background`, `manual`, `off` | `background` | Control qmd update after writes |
 | `PI_MEMORY_NO_SEARCH` | `1` | unset | Disable per-turn search injection |
 | `PI_MEMORY_SUMMARIZE_TRANSITIONS` | `1`, `true`, `yes`, `on` | unset | Also summarize lifecycle transitions |
-| `PI_MEMORY_LEARNING` | `off`, `review`, `auto-review` | `review` | Control session and curator daily learning candidate extraction |
+| `PI_MEMORY_LEARNING` | `off`, `review`, `auto-review` | `review` | Control session learning candidate extraction |
 | `PI_MEMORY_LEARNING_MIN_CONFIDENCE` | `low`, `medium`, `high` | `medium` | Minimum extractor confidence to keep |
 | `PI_MEMORY_SKILL_DRAFTS` | `off`, `review` | `review` | Allow curator to propose disabled skill drafts |
 | `PI_MEMORY_AUTO_APPROVE_MEMORY` | `1`, `true`, `yes`, `on` | unset | YOLO mode for approving newly created memory proposals |
 | `PI_MEMORY_AUTO_APPROVE_SKILL_DRAFTS` | `1`, `true`, `yes`, `on` | unset | YOLO mode for creating newly proposed disabled skill drafts |
-| `PI_EVOLUTION_ENABLED` | `0`, `1`, `true`, `false` | `1` | Enable snapshot + git versioning |
-| `PI_EVOLUTION_DIR` | path | `~/.pi/agent/evolution` | Local evolution repo directory |
-| `PI_EVOLUTION_REMOTE` | URL | unset | Optional personal private Git remote for manual/optional push |
-| `PI_EVOLUTION_BRANCH` | branch | `main` | Local branch used for init/clone |
-| `PI_EVOLUTION_AUTO_COMMIT` | `0`, `1`, `true`, `false` | `1` | Commit sync/snapshot changes automatically |
-| `PI_EVOLUTION_AUTO_PUSH` | `0`, `1`, `true`, `false` | `0` | Push after commits automatically |
-| `PI_EVOLUTION_MAX_SNAPSHOTS` | positive integer | `100` | Maximum snapshots to keep in the local sliding window |
 
 ## Development