(MOT-4018) feat(memory): durable cross-session agent memory worker + console page

.github/scripts/parse_publish_workers_input.py

@@ -20,6 +20,7 @@
     "iii-directory",
     "image-resize",
     "mcp",
+    "memory",
     "scrapling",
     "shell",
     "storage",

.github/workflows/create-tag.yml

@@ -31,6 +31,8 @@ on:
           - image-resize
           - llm-router
           - mcp
+          - memory
+          - memory-consolidate
           - opencode
           - pi
           - provider-anthropic

.github/workflows/release.yml

@@ -25,6 +25,8 @@ on:
       - 'image-resize/v*'
       - 'llm-router/v*'
       - 'mcp/v*'
+      - 'memory/v*'
+      - 'memory-consolidate/v*'
       - 'opencode/v*'
       - 'pi/v*'
       - 'provider-anthropic/v*'

README.md

@@ -57,12 +57,14 @@ npx skills add iii-hq/iii --all
 | [`lsp`](lsp/) | Rust | Language Server for iii function ids, trigger configs, and worker discovery. Autocomplete / hover across JS/TS, Python, Rust. |
 | [`lsp-vscode`](lsp-vscode/) | Node | VS Code extension package `iii-lsp`, embedding the `lsp` server. |
 | [`image-resize`](image-resize/) | Rust | Image resize via channel I/O — JPEG/PNG/WebP with EXIF auto-orient, scale-to-fit / crop-to-fit. |
-| [`llm-router`](llm-router/) | Rust | One front door + provider protocol in front of every LLM provider — `router::chat`/`router::complete`, provider registry + credentials, model catalog, and routing. See [`llm-router/README.md`](llm-router/README.md). |
+| [`llm-router`](llm-router/) | Rust | One front door + provider protocol in front of every LLM provider — `router::chat`/`router::complete`/`router::embed`, provider registry + credentials, model catalog, and routing. See [`llm-router/README.md`](llm-router/README.md). |
 | [`mcp`](mcp/) | Rust | MCP 2025-06-18 Streamable HTTP bridge — exposes iii functions tagged `mcp.expose` as MCP tools. |
+| [`memory`](memory/) | Rust | Durable cross-session agent memory — named banks of always-injected markdown rules and auto-extracted memories, hybrid BM25 + entity + semantic recall, pinning, supersede-never-delete history, and two live trigger types. Plain files on disk; binds the harness `pre-generate` hook for injection and `turn-completed` for background capture. |
+| [`memory-consolidate`](memory-consolidate/) | Rust | Scheduled hygiene sibling of `memory` — deterministic dedup of near-duplicate memories, supersede-only through the public memory functions, pinned untouchable, catch-up-on-boot scheduling. Removable without touching stored memory. |
 | [`rbac-proxy`](rbac-proxy/) | Rust | RBAC boundary proxy for the iii worker protocol — opens its own port and reverse-proxies functions + channels to a trusted engine listener, authenticating each connection, gating every invocation and trigger binding, namespacing registrations, running middleware + registration hooks, and filtering the `engine::*` discovery results to the caller's boundaries. The [`console`](console/) reverse-proxy with the engine's RBAC vendored in front, out of process. |
 | [`provider-anthropic`](provider-anthropic/) | Rust | Anthropic Messages API provider behind `llm-router` — `provider::anthropic::stream` with prompt caching, thinking, and live model discovery. |
 | [`provider-llamacpp`](provider-llamacpp/) | Rust | llama.cpp server (`llama-server`) Chat Completions provider behind `llm-router` — `provider::llamacpp::stream` with optional (no-`--api-key`) auth, real json_schema-constrained output, and live model discovery via `/v1/models` + `/props`. |
-| [`provider-openai`](provider-openai/) | Rust | OpenAI Chat Completions provider behind `llm-router` — `provider::openai::stream` with reasoning support and live chat-model discovery. |
+| [`provider-openai`](provider-openai/) | Rust | OpenAI Chat Completions provider behind `llm-router` — `provider::openai::stream` with reasoning support and live chat-model discovery, plus `provider::openai::embed` for batch embeddings (OpenAI-compatible endpoints included). |
 | [`provider-xai`](provider-xai/) | Rust | xAI (Grok) Chat Completions provider behind `llm-router` — `provider::xai::stream` with grok reasoning support and live model discovery against `api.x.ai`. |
 | [`provider-zai`](provider-zai/) | Rust | Z.AI (GLM) Chat Completions provider behind `llm-router` — `provider::zai::stream` with GLM thinking/effort support and a curated catalog against `api.z.ai` (no upstream model listing). |
 | [`shell`](shell/) | Rust | Unix shell + filesystem worker — `shell::exec` with denylist/timeout/output caps and background jobs; `fs::ls`/`stat`/`mkdir`/`rm`/`chmod`/`mv`/`grep`/`sed`/`read`/`write` with host jail, denylist, and size caps. |

console/Cargo.toml

@@ -2,7 +2,7 @@
 
 [package]
 name = "console"
-version = "1.6.4"
+version = "1.6.5"
 edition = "2021"
 publish = false
 

console/README.md

@@ -65,6 +65,18 @@ The human window into the [`worktree`](../worktree) worker: parallel agent check
 - **Working-dir badge** — a conversation rooted in a managed worktree shows branch, short id, dirty `*` / ahead `+n` indicators, and a lifecycle dot instead of the plain path chip; the raw path stays reachable as the tooltip
 - **Live land notices** — `worktree::landed` / `worktree::land-blocked` events surface in the chat as notices (target branch and merged sha, or the block reason and conflicted files) and refresh the badge
 
+### Memory
+
+The human window into the [`memory`](../memory) worker: named banks of always-injected markdown rules and auto-extracted memories. Lives in [`web/src/pages/Memory/`](web/src/pages/Memory). Presence-gated like Worktrees: the page appears only while the memory worker is connected (a direct `#/memory` hit lands on an install notice).
+
+- **Bank rail** — every bank with live memory / pinned / rule counts and inline create
+- **Rules tab (first)** — the bank's markdown rules as in-place editors (save appears only on edit; delete asks to confirm); the agent appends learned standing instructions to the auto-managed `learned` rule as you correct it in chat
+- **Memories tab** — server-paged newest-first list with pin / edit-in-place / tombstone delete, a show-history toggle, and search that runs `memory::recall` (the ranked scorer, not a client filter)
+- **Graph tab** — entity hubs with memories as spokes: level-of-detail for large banks, draggable nodes, wheel zoom/pan, click a node for an inspect card
+- **Preview tab** — the whole turn before it happens: `memory::preview` composes the exact system-prompt rules section and the appended memories (ambient floor + budgets applied) for a hypothetical question, with clickable example questions from the bank's own content
+- **Live** — the page re-reads off `memory::item-changed` / `memory::bank-changed` (poll fallback), so memories appear the moment they're learned
+- **In chat** — a bank picker in the composer (session metadata `memory_bank`) and a memory chip on each assistant reply naming the bank, how many rules and memories fed the turn, and whether recall ran semantic; click it to expand the exact records
+
 ### Live catalogs
 
 The composer's `@`-mentions and the model picker pull from the engine in real time.

console/web/src/App.tsx

@@ -22,6 +22,7 @@ import { buildViewOptions } from '@/lib/nav-options'
 import { cn } from '@/lib/utils'
 import { Browser } from '@/pages/Browser'
 import { Configuration } from '@/pages/Configuration'
+import { Memory } from '@/pages/Memory'
 import { TracesV2 } from '@/pages/TracesV2'
 import { Workers } from '@/pages/Workers'
 import { Worktrees } from '@/pages/Worktrees'
@@ -97,6 +98,8 @@ export function App() {
               <Worktrees />
             ) : view === 'browser' ? (
               <Browser />
+            ) : view === 'memory' ? (
+              <Memory />
             ) : (
               <TracesV2 />
             )}
@@ -126,8 +129,13 @@ function Header({
   // Optional-worker entries appear only while their worker is present; a
   // direct #/worktrees or #/browser hit still lands on that page's install
   // notice.
-  const { worktreeAvailable, browserAvailable } = useConversationsCtx()
-  const viewOptions = buildViewOptions(worktreeAvailable, browserAvailable)
+  const { worktreeAvailable, browserAvailable, memoryAvailable } =
+    useConversationsCtx()
+  const viewOptions = buildViewOptions(
+    worktreeAvailable,
+    browserAvailable,
+    memoryAvailable,
+  )
   const onConsoleSettings = view === 'configuration'
   return (
     <header className="flex items-center justify-between pl-3 pr-6 h-12 border-b border-rule shrink-0">

console/web/src/components/chat/BankPicker.tsx

@@ -0,0 +1,73 @@
+import { useEffect, useState } from 'react'
+import { Select } from '@/components/ui/Select'
+import { listBanks, type MemoryBank } from '@/lib/memory'
+
+/**
+ * In-chat memory bank picker — the "which memory am I using" control.
+ * Writing blogs? Pick the blog bank and its style rules + memories feed
+ * every turn. Switch to coding and a different memory applies; contexts
+ * never bleed. Selecting `auto` defers to the memory worker's configured
+ * default bank. Lives next to the composer as one compact dropdown and
+ * commits through session metadata (`memory_bank`), so it applies to the
+ * NEXT turn immediately, mid-conversation switches included.
+ */
+
+interface BankPickerProps {
+  /** Bank for THIS conversation; null = the worker's default bank. */
+  value: string | null
+  onChange: (next: string | null) => void
+  disabled?: boolean
+}
+
+const AUTO = '(auto)'
+
+export function BankPicker({ value, onChange, disabled }: BankPickerProps) {
+  const [banks, setBanks] = useState<MemoryBank[]>([])
+
+  useEffect(() => {
+    let cancelled = false
+    void listBanks()
+      .then((next) => {
+        if (!cancelled) setBanks(next)
+      })
+      .catch(() => {})
+    return () => {
+      cancelled = true
+    }
+  }, [])
+
+  const known = banks.some((b) => b.name === value)
+  const options = [
+    {
+      value: AUTO,
+      label: 'memory: auto',
+      title: "use the memory worker's default bank",
+    },
+    ...banks.map((b) => ({
+      value: b.name,
+      label: `memory: ${b.name}`,
+      title: `${b.memories} memories · ${b.rules} rules — rules + recalled memories from this bank feed every turn`,
+    })),
+    // A bank set elsewhere (CLI, another client) that we haven't listed.
+    ...(value && !known
+      ? [
+          {
+            value,
+            label: `memory: ${value}`,
+            title: 'set outside this console',
+          },
+        ]
+      : []),
+  ]
+
+  return (
+    <Select<string>
+      value={value ?? AUTO}
+      onChange={(next) => onChange(next === AUTO ? null : next)}
+      disabled={disabled}
+      aria-label="memory bank"
+      className="min-w-[120px] shrink-0"
+      options={options}
+    />
+  )
+}

console/web/src/components/chat/ChatView.tsx

@@ -1705,6 +1705,14 @@ export function ChatView({
             onModelChange={(next) => onUpdateModel(conversation.id, next)}
             showWorkingDir={workingDirEnabled}
             workingDir={conversation.workingDir ?? null}
+            showMemoryBank={
+              backend.id === 'real' &&
+              (conversationsCtx?.memoryAvailable ?? false)
+            }
+            memoryBank={conversation.memoryBank ?? null}
+            onMemoryBankChange={(next) =>
+              conversationsCtx?.setMemoryBank(conversation.id, next)
+            }
             workingDirLocked={false}
             workingDirError={workingDirError}
             defaultWorkingDir={defaultWorkingDir}

console/web/src/components/chat/Composer.tsx

@@ -19,6 +19,7 @@ import type {
 } from '@/types/chat'
 import { AttachmentButton } from './AttachmentButton'
 import { AttachmentChip } from './AttachmentChip'
+import { BankPicker } from './BankPicker'
 import { DirectoryPicker, type WorktreePickerOptions } from './DirectoryPicker'
 import { LexicalShell } from './LexicalShell'
 import { ModelPicker } from './ModelPicker'
@@ -61,6 +62,11 @@ interface ComposerProps {
   /** Show the per-session working-directory picker (real backend only). */
   showWorkingDir?: boolean
   workingDir?: string | null
+  /** Show the memory bank picker (memory worker present, real backend). */
+  showMemoryBank?: boolean
+  /** This chat's memory bank; null = the worker's default bank. */
+  memoryBank?: string | null
+  onMemoryBankChange?: (next: string | null) => void
   /**
    * When true, the picker renders read-only. NOT set after the first send —
    * the working dir stays re-scopable mid-conversation (ChatView always
@@ -140,6 +146,9 @@ export function Composer({
   thinkingLevel,
   showWorkingDir,
   workingDir,
+  showMemoryBank,
+  memoryBank,
+  onMemoryBankChange,
   workingDirLocked,
   workingDirError,
   defaultWorkingDir,
@@ -339,6 +348,13 @@ export function Composer({
               className="shrink-0"
             />
           ) : null}
+          {showMemoryBank && onMemoryBankChange ? (
+            <BankPicker
+              value={memoryBank ?? null}
+              onChange={onMemoryBankChange}
+              disabled={optionsDisabled}
+            />
+          ) : null}
           {showPermissionMode ? (
             <PermissionModePicker
               value={permissionMode}

console/web/src/components/chat/MemoryChip.tsx

@@ -0,0 +1,107 @@
+import { useState } from 'react'
+import { getMemory, type MemoryItem } from '@/lib/memory'
+import { cn } from '@/lib/utils'
+
+/**
+ * The "what fed this reply" chip on assistant messages: bank name + how
+ * many memories the memory worker injected into the generate. Click to
+ * expand the exact memories (fetched by id on first open) — memory acting
+ * visibly inside the conversation, not just in traces.
+ */
+
+interface MemoryChipProps {
+  memory: {
+    bank: string
+    memories: number
+    memoryIds: string[]
+    rules?: number
+    truncated?: boolean
+    semantic?: boolean
+  }
+}
+
+export function MemoryChip({ memory }: MemoryChipProps) {
+  const [open, setOpen] = useState(false)
+  const [memories, setMemories] = useState<MemoryItem[] | null>(null)
+  const [loading, setLoading] = useState(false)
+  const [loadError, setLoadError] = useState(false)
+
+  const parts = []
+  if (memory.memories > 0)
+    parts.push(`${memory.memories} memor${memory.memories === 1 ? 'y' : 'ies'}`)
+  if (memory.rules)
+    parts.push(`${memory.rules} rule${memory.rules === 1 ? '' : 's'}`)
+  if (parts.length === 0) return null
+
+  const toggle = async () => {
+    const next = !open
+    setOpen(next)
+    if (next && memories === null && memory.memoryIds.length > 0) {
+      setLoading(true)
+      setLoadError(false)
+      try {
+        const loaded = await Promise.all(
+          memory.memoryIds.map((id) => getMemory(memory.bank, id)),
+        )
+        setMemories(loaded.filter((m): m is MemoryItem => m !== null))
+      } catch {
+        // A failed fetch is not "no memories": show it as an error and
+        // let the next open retry.
+        setLoadError(true)
+      } finally {
+        setLoading(false)
+      }
+    }
+  }
+
+  return (
+    <span className="inline-flex flex-col items-start normal-case tracking-normal">
+      <button
+        type="button"
+        onClick={() => void toggle()}
+        title="what the memory worker fed this reply — click for the exact memories"
+        className={cn(
+          'font-mono text-[10px] lowercase px-1.5 py-0.5 border transition-colors',
+          open
+            ? 'border-accent text-ink'
+            : 'border-rule text-ink-faint hover:border-ink hover:text-ink',
+        )}
+      >
+        memory: {memory.bank} · {parts.join(' · ')}
+        {memory.semantic ? ' · semantic' : ''}
+        {memory.truncated ? ' · truncated' : ''}
+      </button>
+      {open ? (
+        <span className="mt-1 flex flex-col gap-1 border border-rule-2 bg-panel px-2 py-1.5 max-w-md">
+          {loading ? (
+            <span className="font-mono text-[10px] lowercase text-ink-ghost">
+              loading memories…
+            </span>
+          ) : loadError ? (
+            <span className="font-mono text-[10px] lowercase text-alert">
+              could not load memories — click to retry
+            </span>
+          ) : memories && memories.length > 0 ? (
+            memories.map((memory) => (
+              <span
+                key={memory.id}
+                className="font-mono text-[11px] text-ink leading-snug"
+              >
+                - {memory.text}
+                {memory.pinned ? (
+                  <span className="text-accent"> ·pinned</span>
+                ) : null}
+              </span>
+            ))
+          ) : (
+            <span className="font-mono text-[10px] lowercase text-ink-ghost">
+              {memory.memoryIds.length === 0
+                ? 'only rules fed this turn (always-injected markdown)'
+                : 'memories no longer available (superseded or bank changed)'}
+            </span>
+          )}
+        </span>
+      ) : null}
+    </span>
+  )
+}

console/web/src/components/chat/Message.tsx

@@ -12,6 +12,7 @@ import type {
   UserMessage as UserMessageType,
 } from '@/types/chat'
 import { AttachmentChip } from './AttachmentChip'
+import { MemoryChip } from './MemoryChip'
 import { ThoughtMessage } from './ThoughtMessage'
 
 interface MessageProps {
@@ -292,14 +293,15 @@ function AssistantMessage({ message }: { message: AssistantMessageType }) {
   const showCaret = !!message.streaming
   return (
     <article className="flex flex-col gap-2">
-      <header className="font-mono text-[11px] uppercase tracking-[0.06em] text-ink-ghost flex items-center gap-2">
+      <header className="font-mono text-[11px] uppercase tracking-[0.06em] text-ink-ghost flex items-center gap-2 flex-wrap">
         <Prompt symbol=">">assistant</Prompt>
         {message.model ? (
           <span className="text-ink-ghost">· {message.model}</span>
         ) : null}
         {message.mode ? (
           <span className="text-ink-ghost">· {message.mode}</span>
         ) : null}
+        {message.memory ? <MemoryChip memory={message.memory} /> : null}
       </header>
       <div className="pr-1">
         {message.content ? (