opencode-claude-memory 1.7.7 → 2.0.0

package/dist/index.js

@@ -1,675 +1,60 @@
-import { tool } from "@opencode-ai/plugin";
-import { parse, resolve } from "path";
-import { buildMemorySystemPrompt } from "./prompt.js";
-import { formatRecalledMemories, recallSelectedMemories } from "./recall.js";
-import { isSupportedRecallSelectorClient, selectRelevantMemoryFilenames } from "./recallSelector.js";
-import { scanMemoryFiles } from "./memoryScan.js";
-import { saveMemoryDetailed, deleteMemory, listMemories, searchMemories, readMemory, MEMORY_TYPES, } from "./memory.js";
-import { getMemoryDir } from "./paths.js";
-import { formatMemorySaveResult, getNativeExtractMaxSteps, getNativeExtractTimeoutMs, logNativeExtractionFailure, } from "./nativeExtraction.js";
-const turnContextBySession = new Map();
-const selectorSessionIDs = new Set();
-function shouldIgnoreMemoryContext(query) {
-    if (process.env.OPENCODE_MEMORY_IGNORE === "1")
-        return true;
-    if (!query)
-        return false;
-    const normalized = query.toLowerCase();
-    return (/(ignore|don't use|do not use|without|skip)\s+(the\s+)?memory/.test(normalized) ||
-        /memory\s+(should be|must be)?\s*ignored/.test(normalized));
-}
-function extractUserQuery(message) {
-    if (!message || typeof message !== "object")
-        return undefined;
-    if ("content" in message) {
-        const content = message.content;
-        if (typeof content === "string")
-            return content;
-        if (content !== undefined)
-            return JSON.stringify(content);
-    }
-    if ("parts" in message) {
-        const parts = message.parts;
-        if (Array.isArray(parts)) {
-            const text = parts
-                .map((part) => {
-                if (!part || typeof part !== "object")
-                    return "";
-                return typeof part.text === "string"
-                    ? part.text
-                    : "";
-            })
-                .filter(Boolean)
-                .join("\n")
-                .trim();
-            if (text)
-                return text;
-        }
-    }
-    return undefined;
-}
-function getLastUserQuery(messages) {
-    for (let i = messages.length - 1; i >= 0; i--) {
-        const message = messages[i];
-        if (message?.info?.role !== "user")
-            continue;
-        const query = extractUserQuery(message);
-        const sessionID = typeof message.info?.sessionID === "string" ? message.info.sessionID : undefined;
-        const messageID = typeof message.info?.id === "string" ? message.info.id : undefined;
-        return { query, sessionID, messageID, messageIndex: i };
-    }
-    return {};
-}
-function isAutoMemoryPart(part) {
-    if (!part || typeof part !== "object")
-        return false;
-    return typeof part.text === "string" &&
-        part.text.includes("# Auto Memory");
-}
-// Parses "### <name> (<type>)" headers from the ## Recalled Memories section
-// of system prompts. After compaction old system messages disappear, so
-// the returned set naturally shrinks — no manual reset needed.
-function extractSurfacedMemoryKeys(systemText) {
-    const keys = new Set();
-    const recalledSection = systemText.indexOf("## Recalled Memories");
-    if (recalledSection === -1)
-        return keys;
-    const headerPattern = /^### (.+?) \((\w+)\)/gm;
-    const section = systemText.slice(recalledSection);
-    for (let match = headerPattern.exec(section); match !== null; match = headerPattern.exec(section)) {
-        keys.add(`${match[1]}|${match[2]}`);
-    }
-    return keys;
-}
-// Only completed tools — matches Claude Code's collectRecentSuccessfulTools().
-function extractRecentTools(messages) {
-    const tools = [];
-    const seen = new Set();
-    for (const message of messages) {
-        if (!message.parts || !Array.isArray(message.parts))
-            continue;
-        for (const part of message.parts) {
-            if (!part || typeof part !== "object")
-                continue;
-            const p = part;
-            if (p.type !== "tool" || !p.tool)
-                continue;
-            if (p.state?.status !== "completed")
-                continue;
-            if (seen.has(p.tool))
-                continue;
-            seen.add(p.tool);
-            tools.push(p.tool);
-        }
-    }
-    return tools;
-}
-function getRecallAgent() {
-    return process.env.OPENCODE_MEMORY_RECALL_AGENT || "opencode-memory-recall";
-}
-function getRecallModel() {
-    const raw = process.env.OPENCODE_MEMORY_RECALL_MODEL;
-    if (!raw)
-        return undefined;
-    const slashIdx = raw.indexOf("/");
-    if (slashIdx <= 0 || slashIdx === raw.length - 1)
-        return undefined;
+import { AgentRegistry } from "./agents.js";
+import { parseConfig } from "./config.js";
+import { ExtractionCoordinator } from "./extraction/ExtractionCoordinator.js";
+import { buildMemorySystemPrompt } from "./prompt/systemPrompt.js";
+import { formatRecalledMemories } from "./recall/format.js";
+import { RecallCoordinator } from "./recall/RecallCoordinator.js";
+import { MemoryStore } from "./store/MemoryStore.js";
+import { resolveMemoryRoot } from "./store/paths.js";
+import { buildMemoryTools } from "./tools.js";
+import { createLogger } from "./util/log.js";
+import { OwnedSessions } from "./util/ownedSessions.js";
+export const PLUGIN_ID = "opencode-claude-memory";
+// Assembly only. Every piece of mutable state lives on the coordinators created here, so OpenCode's
+// multi-directory `serve` gets fully isolated instances. `env` and `homeDir` are injectable so tests
+// never touch the real environment (CLAUDE_CONFIG_DIR is the only variable read, see config.ts).
+export const createMemoryPlugin = (env, homeDir) => async ({ worktree, directory, client }, options) => {
+    const config = parseConfig(options, env, homeDir);
+    const dir = directory ?? worktree;
+    const store = new MemoryStore(resolveMemoryRoot(worktree, dir), config);
+    const log = createLogger(client, dir);
+    const owned = new OwnedSessions();
+    const agents = new AgentRegistry(config.agents);
+    const deps = { store, config, client, directory: dir, owned, agents, log };
+    const recall = new RecallCoordinator(deps);
+    const extraction = new ExtractionCoordinator(deps);
     return {
-        providerID: raw.slice(0, slashIdx),
-        modelID: raw.slice(slashIdx + 1),
-    };
-}
-function isRootPath(path) {
-    const resolved = resolve(path);
-    return resolved === parse(resolved).root;
-}
-function resolveMemoryRoot(worktree, directory) {
-    if (isRootPath(worktree) && !isRootPath(directory))
-        return directory;
-    return worktree;
-}
-function isUsefulRecallQuery(query) {
-    const trimmed = query?.trim();
-    if (!trimmed)
-        return false;
-    if (/\s/.test(trimmed))
-        return true;
-    return /[\u3400-\u9fff]/.test(trimmed) && trimmed.length >= 4;
-}
-function buildTurnID(sessionID, messageID, messageIndex, query) {
-    return `${sessionID}:${messageID ?? `${messageIndex ?? -1}:${query ?? ""}`}`;
-}
-function alreadySurfacedKey(header) {
-    return `${header.name ?? header.filename.replace(/\.md$/, "").replace(/.*\//, "")}|${header.type ?? "user"}`;
-}
-function startRecallPrefetch(input) {
-    if (!input.client || !isUsefulRecallQuery(input.query))
-        return undefined;
-    if (!isSupportedRecallSelectorClient(input.client))
-        return undefined;
-    const memoryDir = getMemoryDir(input.worktree);
-    const headers = scanMemoryFiles(memoryDir).filter((header) => !input.alreadySurfaced.has(alreadySurfacedKey(header)));
-    if (headers.length === 0)
-        return undefined;
-    const handle = {
-        turnID: input.turnID,
-        settled: false,
-        consumed: false,
-        result: [],
-    };
-    const promise = selectRelevantMemoryFilenames({
-        client: input.client,
-        directory: input.directory,
-        parentSessionID: input.parentSessionID,
-        query: input.query,
-        memories: headers,
-        recentTools: input.recentTools,
-        selectorSessionIDs,
-        agent: getRecallAgent(),
-        model: getRecallModel(),
-    })
-        .then((selectedFilenames) => recallSelectedMemories(headers, selectedFilenames, input.alreadySurfaced))
-        .catch(() => []);
-    void promise.then((result) => {
-        handle.result = result;
-    }).finally(() => {
-        handle.settled = true;
-    });
-    return handle;
-}
-function consumeRecallPrefetch(ctx) {
-    const prefetch = ctx?.recallPrefetch;
-    if (!prefetch || !prefetch.settled || prefetch.consumed)
-        return [];
-    prefetch.consumed = true;
-    return prefetch.result;
-}
-// Tracks how many memory entries a memory_list call saw so tool.execute.after
-// can render a meaningful title without re-reading the filesystem. Keyed by
-// callID, which uniquely identifies a single tool invocation.
-const memoryListCountByCallID = new Map();
-const memorySearchCountByCallID = new Map();
-function buildMemoryToolTitle(toolID, args, callID) {
-    switch (toolID) {
-        case "memory_save": {
-            const type = typeof args?.type === "string" ? args.type : "";
-            const name = typeof args?.name === "string" ? args.name : "";
-            if (type && name)
-                return `${type}: ${name}`;
-            if (name)
-                return name;
-            return undefined;
-        }
-        case "memory_delete":
-        case "memory_read": {
-            const fileName = typeof args?.file_name === "string" ? args.file_name : "";
-            return fileName || undefined;
-        }
-        case "memory_list": {
-            const count = callID ? memoryListCountByCallID.get(callID) : undefined;
-            if (callID)
-                memoryListCountByCallID.delete(callID);
-            if (count === undefined)
-                return "list memories";
-            return `${count} ${count === 1 ? "memory" : "memories"}`;
-        }
-        case "memory_search": {
-            const query = typeof args?.query === "string" ? args.query : "";
-            const count = callID ? memorySearchCountByCallID.get(callID) : undefined;
-            if (callID)
-                memorySearchCountByCallID.delete(callID);
-            if (query && count !== undefined) {
-                return `"${query}" · ${count} ${count === 1 ? "match" : "matches"}`;
-            }
-            if (query)
-                return `"${query}"`;
-            return undefined;
-        }
-        default:
-            return undefined;
-    }
-}
-function getCallID(ctx) {
-    if (!ctx || typeof ctx !== "object")
-        return undefined;
-    const v = ctx.callID;
-    return typeof v === "string" ? v : undefined;
-}
-function getSessionID(ctx) {
-    if (!ctx || typeof ctx !== "object")
-        return undefined;
-    const v = ctx.sessionID;
-    return typeof v === "string" ? v : undefined;
-}
-// ─── Native post-session extraction (optional) ───────────────────────────────
-// Architecture follows opencode-mem's proven pattern: hook `session.idle`,
-// fetch the conversation via ctx.client.session.messages(), then run extraction
-// in an isolated sub-session (create + prompt + delete via ctx.client). No shell,
-// no `opencode run --fork` subprocess — those don't work cross-platform from a
-// plugin and `ctx.$` is not reliably populated.
-// Enabled by default on Windows (bash wrapper can't run there); opt-in elsewhere
-// via OPENCODE_MEMORY_NATIVE_EXTRACT=1; =0 force-disables.
-// 10s debounce collapses rapid idle events. Recursion guard via an in-process
-// Set of sub-session IDs. Extraction is additive only (memory_save).
-const EXTRACT_PROMPT = `You are now acting as the memory extraction subagent. The conversation below is reviewed for anything worth remembering for future sessions.
-
-## What to save
-
-Use the \`memory_save\` tool to persist memories. There are four types:
-
-1. **user** — Who the user is: role, expertise, preferences, communication style. Helps tailor future interactions.
-2. **feedback** — Guidance on how to work: corrections ("don't do X"), confirmations ("yes, keep doing that"), approach preferences. Include *why* so edge cases can be judged.
-3. **project** — Ongoing work context: goals, deadlines, initiatives, decisions, bugs. NOT derivable from code/git. Convert relative dates to absolute.
-4. **reference** — Pointers to external resources: URLs, tool names, where to find information outside the codebase.
-
-## What NOT to save
-
-- Code patterns, architecture, file structure — derivable from the codebase
-- Git history, recent changes — use \`git log\`/\`git blame\`
-- Debugging solutions — the fix is in the code
-- Anything already in AGENTS.md / project config files
-- Ephemeral task details or current conversation context
-- Information that was already saved in a previous extraction
-
-## How to save
-
-For each memory worth saving, call \`memory_save\` with:
-- \`file_name\`: descriptive slug (e.g., \`user_role\`, \`feedback_testing_approach\`)
-- \`name\`: short title
-- \`description\`: one-line description (used for relevance matching in future sessions)
-- \`type\`: one of user, feedback, project, reference
-- \`content\`: the memory content. For feedback/project types, structure as: rule/fact, then **Why:** and **How to apply:** lines.
-
-## Instructions
-
-1. Analyze the conversation for memorable information
-2. Check existing memories first (use \`memory_list\`) to avoid duplicates — update existing ones if needed
-3. Save each distinct memory as a separate entry
-4. If the conversation was trivial (e.g., just "hello" or a quick lookup), save nothing — that's fine
-5. Be selective: 0-3 memories per session is typical. Quality over quantity.
-6. Do NOT save a memory about the extraction process itself.
-7. Each \`memory_save\` call persists immediately and its result lists everything saved so far in this run. Never save the same \`file_name\` twice in one run. When nothing else is worth saving, stop calling tools and reply with a one-line summary.`;
-const NATIVE_EXTRACT_DEBOUNCE_MS = 10000;
-const NATIVE_EXTRACT_MAX_CONV_CHARS = 60000;
-const NATIVE_EXTRACT_GRACE_MS = 60000; // keep forkID in the guard after delete — covers the idle-race window
-const nativeIdleTimer = new Map();
-const nativeExtractionSessions = new Set(); // sub-session IDs → skip their idle AND shield from transforms
-const nativeExtractionInFlight = new Set(); // parent sessionIDs with an extraction running → no overlap
-const nativeExtractionSavedFiles = new Map(); // fork sessionID → file names memory_save wrote this run (#35)
-function extractIDFromResponse(response) {
-    const data = response?.data ?? response;
-    const obj = data;
-    const v = obj?.id ?? obj?.sessionID;
-    return typeof v === "string" ? v : null;
-}
-function getNativeExtractAgent() {
-    return process.env.OPENCODE_MEMORY_AGENT || "opencode-memory-extract";
-}
-// Parse OPENCODE_MEMORY_MODEL into the SDK's ModelRef shape (providerID/modelID), same as getRecallModel().
-function getNativeExtractModel() {
-    const raw = process.env.OPENCODE_MEMORY_MODEL;
-    if (!raw)
-        return undefined;
-    const slashIdx = raw.indexOf("/");
-    if (slashIdx <= 0 || slashIdx === raw.length - 1)
-        return undefined;
-    return { providerID: raw.slice(0, slashIdx), modelID: raw.slice(slashIdx + 1) };
-}
-function buildConversationForExtraction(messages) {
-    const lines = [];
-    for (const m of messages) {
-        const role = m?.info?.role;
-        if (!role || !Array.isArray(m?.parts))
-            continue;
-        for (const p of m.parts) {
-            const part = p;
-            if (part.type === "text" && typeof part.text === "string" && !part.synthetic) {
-                lines.push(`### ${role === "user" ? "User" : "Assistant"}\n${part.text}`);
-            }
-            else if (part.type === "tool" && part.tool) {
-                if (part.state?.status === "completed" && typeof part.state?.output === "string") {
-                    const out = part.state.output.length > 300 ? part.state.output.slice(0, 300) + "…" : part.state.output;
-                    lines.push(`_[tool ${part.tool}: ${out}]_`);
-                }
-            }
-        }
-    }
-    let text = lines.join("\n\n");
-    // Keep the TAIL (newest turns carry the new facts worth extracting), drop the oldest head.
-    if (text.length > NATIVE_EXTRACT_MAX_CONV_CHARS) {
-        text = "…[older turns truncated]\n\n" + text.slice(-NATIVE_EXTRACT_MAX_CONV_CHARS);
-    }
-    return text;
-}
-// Security: the extraction fork runs on raw, potentially-untrusted transcript content (fetched web
-// pages, tool output). It MUST be sandboxed to the memory tools only (no bash/edit/write) and capped
-// by a timeout so a hung prompt stops blocking best-effort sub-session cleanup.
-async function runNativeExtraction(client, sessionID, directory) {
-    const c = client;
-    if (!c?.session?.messages || !c.session.create || !c.session.prompt)
-        return;
-    if (nativeExtractionInFlight.has(sessionID))
-        return; // no overlapping runs against the same session
-    nativeExtractionInFlight.add(sessionID);
-    let forkID = null;
-    let promptSettled = false;
-    try {
-        // query.directory is required for correct resolution under multi-directory serve.
-        const resp = await c.session.messages({ path: { id: sessionID }, query: { directory } });
-        const messages = (resp.data ?? []);
-        const convText = buildConversationForExtraction(messages);
-        if (convText.trim().length < 20)
-            return;
-        const created = await c.session.create({
-            body: { parentID: sessionID, title: "opencode-memory extraction" },
-            query: { directory },
-        });
-        forkID = extractIDFromResponse(created);
-        if (!forkID)
-            return;
-        nativeExtractionSessions.add(forkID); // shield the fork from the plugin's own recall/transform hooks
-        const body = {
-            agent: getNativeExtractAgent(),
-            system: EXTRACT_PROMPT, // always pass explicitly — ??= agent registration can miss if the agent already exists
-            tools: { "*": false, memory_save: true, memory_list: true }, // wildcard deny first; findLast → memory allows win
-            parts: [{ type: "text", text: convText }],
-        };
-        const extractModel = getNativeExtractModel();
-        if (extractModel)
-            body.model = extractModel;
-        // Race the prompt against a timeout so a hung/permission-gated fork stops blocking cleanup.
-        let timer;
-        const timeout = new Promise((_, reject) => {
-            timer = setTimeout(() => reject(new Error("native extraction timed out")), getNativeExtractTimeoutMs());
-        });
-        try {
-            await Promise.race([
-                c.session.prompt({ path: { id: forkID }, query: { directory }, body }),
-                timeout,
-            ]);
-            promptSettled = true;
-        }
-        finally {
-            if (timer)
-                clearTimeout(timer);
-        }
-    }
-    catch (e) {
-        logNativeExtractionFailure(client, directory, sessionID, e);
-    }
-    finally {
-        if (forkID) {
-            // On timeout the server is still running the fork; stop it before deleting the session so it
-            // does not keep inserting parts/messages for a row that no longer exists (FK constraint errors).
-            if (!promptSettled) {
-                await c.session?.abort?.({ path: { id: forkID }, query: { directory } }).catch(() => { });
-            }
-            await c.session?.delete?.({ path: { id: forkID }, query: { directory } }).catch(() => { });
-            // Hold the guard past delete: the fork's session.idle can arrive on a separate channel after
-            // the delete HTTP call resolves, and would otherwise re-trigger extraction of the fork itself.
-            const id = forkID;
-            setTimeout(() => {
-                nativeExtractionSessions.delete(id);
-                nativeExtractionSavedFiles.delete(id);
-            }, NATIVE_EXTRACT_GRACE_MS);
-        }
-        nativeExtractionInFlight.delete(sessionID);
-    }
-}
-// Native extraction is the bash-wrapper replacement. Default-on for Windows (where the wrapper can't
-// run); opt-in elsewhere. Honors the wrapper's documented OPENCODE_MEMORY_EXTRACT=0 opt-out so a user
-// who disabled extraction gets it disabled here too (not silently re-enabled by the win32 default).
-// OPENCODE_MEMORY_NATIVE_EXTRACT=0/1 force-disables/enables regardless of platform.
-function nativeExtractionEnabled() {
-    if (process.env.OPENCODE_MEMORY_EXTRACT === "0")
-        return false;
-    const flag = process.env.OPENCODE_MEMORY_NATIVE_EXTRACT;
-    if (flag === "0")
-        return false;
-    if (flag === "1")
-        return true;
-    return process.platform === "win32";
-}
-export const MemoryPlugin = async ({ worktree, directory, client }) => {
-    directory ??= worktree;
-    const memoryRoot = resolveMemoryRoot(worktree, directory);
-    getMemoryDir(memoryRoot);
-    return {
-        config: async (config) => {
-            const agentName = getRecallAgent();
-            const mutable = config;
-            mutable.agent ??= {};
-            mutable.agent[agentName] ??= {
-                mode: "all",
-                hidden: true,
-                prompt: "Select up to 5 relevant memory filenames for the current user query. Return only the requested structured output.",
-            };
-            // Dedicated hidden agent for native extraction: its system prompt IS the extraction instruction,
-            // so the fork runs with an explicit system (not the build agent's) and tool-restricted to memory_*.
-            // `steps` caps the fork's agentic iterations so a model that keeps re-saving the same memories
-            // terminates cleanly (OpenCode forces a text-only reply) instead of spinning until the timeout (#35).
-            const extractAgent = (mutable.agent[getNativeExtractAgent()] ??= {
-                mode: "all",
-                hidden: true,
-                prompt: EXTRACT_PROMPT,
-            });
-            const maxSteps = getNativeExtractMaxSteps();
-            if (maxSteps !== undefined && extractAgent.steps === undefined && extractAgent.maxSteps === undefined) {
-                extractAgent.steps = maxSteps;
-            }
+        config: async (cfg) => {
+            agents.register(cfg);
+            // Runs once the agent sandbox is known; failures are logged inside catchUp().
+            void extraction.catchUp();
         },
-        // Native post-session extraction. Enabled by default on Windows (where the
-        // bash wrapper can't run) and via OPENCODE_MEMORY_NATIVE_EXTRACT=1 elsewhere;
-        // =0 force-disables. Hooks `session.idle` (the established idle signal — used
-        // by opencode-mem and others). 10s debounce collapses rapid idle into one
-        // capture; an in-process Set of sub-session IDs prevents recursion. All work
-        // goes through ctx.client — no shell, no fork.
-        event: async (input) => {
-            if (!nativeExtractionEnabled())
-                return;
-            const evt = (input && typeof input === "object" && "event" in input
-                ? input.event
-                : input);
-            if (evt?.type !== "session.idle")
-                return;
-            const sessionID = evt.properties?.sessionID;
-            if (!sessionID)
-                return;
-            if (nativeExtractionSessions.has(sessionID))
-                return; // skip our own extraction sub-sessions
-            if (selectorSessionIDs.has(sessionID))
-                return; // skip recall-selector child sessions (same guard the transforms use)
-            if (nativeIdleTimer.has(sessionID))
-                clearTimeout(nativeIdleTimer.get(sessionID));
-            nativeIdleTimer.set(sessionID, setTimeout(() => {
-                nativeIdleTimer.delete(sessionID);
-                void runNativeExtraction(client, sessionID, directory).catch((e) => {
-                    logNativeExtractionFailure(client, directory, sessionID, e);
-                });
-            }, NATIVE_EXTRACT_DEBOUNCE_MS));
-        },
-        "chat.params": async (input, output) => {
-            if (input.agent !== getRecallAgent())
-                return;
-            output.temperature = 0;
-            output.options = {
-                ...output.options,
-            };
-        },
-        "tool.execute.after": async (input, output) => {
-            if (!input.tool.startsWith("memory_"))
-                return;
-            const title = buildMemoryToolTitle(input.tool, input.args, input.callID);
-            if (title)
-                output.title = title;
+        event: async ({ event }) => {
+            if (event.type === "session.deleted")
+                owned.release(event.properties.info.id, 5_000);
+            recall.onEvent(event);
+            extraction.onEvent(event);
         },
         "experimental.chat.messages.transform": async (_input, output) => {
-            const { query, sessionID, messageID, messageIndex } = getLastUserQuery(output.messages);
-            if (sessionID && (selectorSessionIDs.has(sessionID) || nativeExtractionSessions.has(sessionID)))
-                return;
-            if (sessionID) {
-                const alreadySurfaced = new Set();
-                for (const message of output.messages) {
-                    const role = String(message.info.role);
-                    if (role !== "system")
-                        continue;
-                    for (const part of message.parts) {
-                        if (!part || typeof part !== "object")
-                            continue;
-                        const text = part.text;
-                        if (typeof text === "string") {
-                            for (const key of extractSurfacedMemoryKeys(text)) {
-                                alreadySurfaced.add(key);
-                            }
-                        }
-                    }
-                }
-                const recentTools = extractRecentTools(output.messages);
-                const turnID = buildTurnID(sessionID, messageID, messageIndex, query);
-                const existing = turnContextBySession.get(sessionID);
-                const ignoreMemoryContext = process.env.OPENCODE_MEMORY_IGNORE === "1" || shouldIgnoreMemoryContext(query);
-                let recallPrefetch;
-                if (!ignoreMemoryContext) {
-                    recallPrefetch = existing?.turnID === turnID
-                        ? existing.recallPrefetch
-                        : startRecallPrefetch({
-                            client: client,
-                            directory,
-                            worktree: memoryRoot,
-                            parentSessionID: sessionID,
-                            turnID,
-                            query,
-                            alreadySurfaced,
-                            recentTools,
-                        });
-                }
-                turnContextBySession.set(sessionID, { turnID, query, alreadySurfaced, recentTools, recallPrefetch });
-            }
-            if (shouldIgnoreMemoryContext(query)) {
-                output.messages = output.messages
-                    .map((message) => {
-                    const role = String(message.info.role);
-                    if (role !== "system")
-                        return message;
-                    const parts = message.parts.filter((part) => !isAutoMemoryPart(part));
-                    return { ...message, parts };
-                })
-                    .filter((message) => message.parts.length > 0);
-            }
+            recall.onMessagesTransform(output);
         },
-        "experimental.chat.system.transform": async (_input, output) => {
-            let sessionID;
-            if (_input && typeof _input === "object") {
-                sessionID = typeof _input.sessionID === "string"
-                    ? _input.sessionID
-                    : undefined;
-            }
-            if (sessionID && (selectorSessionIDs.has(sessionID) || nativeExtractionSessions.has(sessionID)))
+        "experimental.chat.system.transform": async ({ sessionID }, output) => {
+            if (owned.has(sessionID))
                 return;
-            const ctx = sessionID ? turnContextBySession.get(sessionID) : undefined;
-            const query = ctx?.query;
-            const ignoreMemoryContext = process.env.OPENCODE_MEMORY_IGNORE === "1" || shouldIgnoreMemoryContext(query);
-            const recalled = ignoreMemoryContext ? [] : consumeRecallPrefetch(ctx);
-            const recalledSection = formatRecalledMemories(recalled);
-            const memoryPrompt = buildMemorySystemPrompt(memoryRoot, recalledSection, {
-                includeIndex: !ignoreMemoryContext,
-            });
-            output.system.push(memoryPrompt);
+            const recalled = await recall.takeRecalled(sessionID);
+            output.system.push(buildMemorySystemPrompt(store, formatRecalledMemories(recalled), {
+                includeIndex: !recall.isIgnored(sessionID),
+            }));
         },
-        tool: {
-            memory_save: tool({
-                description: "Save or update a memory for future conversations. " +
-                    "Each memory is stored as a markdown file with frontmatter. " +
-                    "Use this when the user explicitly asks you to remember something, " +
-                    "or when you observe important information worth preserving across sessions " +
-                    "(user preferences, feedback, project context, external references). " +
-                    "Check existing memories first with memory_list or memory_search to avoid duplicates.",
-                args: {
-                    file_name: tool.schema
-                        .string()
-                        .describe('File name for the memory (without .md extension). Use snake_case, e.g. "user_role", "feedback_testing_style", "project_auth_rewrite"'),
-                    name: tool.schema.string().describe("Human-readable name for this memory"),
-                    description: tool.schema
-                        .string()
-                        .describe("One-line description — used to decide relevance in future conversations, so be specific"),
-                    type: tool.schema
-                        .enum(MEMORY_TYPES)
-                        .describe("Memory type: user (about the person), feedback (guidance on approach), project (ongoing work context), reference (pointers to external systems)"),
-                    content: tool.schema
-                        .string()
-                        .describe("Memory content. For feedback/project types, structure as: rule/fact, then **Why:** and **How to apply:** lines"),
-                },
-                async execute(args, ctx) {
-                    const outcome = saveMemoryDetailed(memoryRoot, args.file_name, args.name, args.description, args.type, args.content);
-                    const sessionID = getSessionID(ctx);
-                    if (!sessionID || !nativeExtractionSessions.has(sessionID))
-                        return formatMemorySaveResult(outcome);
-                    // Extraction fork: record every save so the tool result can carry a durable done-signal.
-                    let saved = nativeExtractionSavedFiles.get(sessionID);
-                    if (!saved)
-                        nativeExtractionSavedFiles.set(sessionID, (saved = []));
-                    saved.push(outcome.fileName);
-                    return formatMemorySaveResult(outcome, saved);
-                },
-            }),
-            memory_delete: tool({
-                description: "Delete a memory that is outdated, wrong, or no longer relevant. Also removes it from the index.",
-                args: {
-                    file_name: tool.schema.string().describe("File name of the memory to delete (with or without .md extension)"),
-                },
-                async execute(args, _ctx) {
-                    const deleted = deleteMemory(memoryRoot, args.file_name);
-                    return deleted ? `Memory "${args.file_name}" deleted.` : `Memory "${args.file_name}" not found.`;
-                },
-            }),
-            memory_list: tool({
-                description: "List all saved memories with their names, types, and descriptions. " +
-                    "Use this to check what memories exist before saving a new one (to avoid duplicates) " +
-                    "or when you need to recall what's been stored.",
-                args: {},
-                async execute(_args, ctx) {
-                    const entries = listMemories(memoryRoot);
-                    const callID = getCallID(ctx);
-                    if (callID)
-                        memoryListCountByCallID.set(callID, entries.length);
-                    if (entries.length === 0) {
-                        return "No memories saved yet.";
-                    }
-                    const lines = entries.map((e) => `- **${e.name}** (${e.type}) [${e.fileName}]: ${e.description}`);
-                    return `${entries.length} memories found:\n${lines.join("\n")}`;
-                },
-            }),
-            memory_search: tool({
-                description: "Search memories by keyword. Searches across names, descriptions, and content. " +
-                    "Use this to find relevant memories before answering questions or when the user references past conversations.",
-                args: {
-                    query: tool.schema.string().describe("Search query — searches across name, description, and content"),
-                },
-                async execute(args, ctx) {
-                    const results = searchMemories(memoryRoot, args.query);
-                    const callID = getCallID(ctx);
-                    if (callID)
-                        memorySearchCountByCallID.set(callID, results.length);
-                    if (results.length === 0) {
-                        return `No memories matching "${args.query}".`;
-                    }
-                    const lines = results.map((e) => `- **${e.name}** (${e.type}) [${e.fileName}]: ${e.description}\n  Content: ${e.content.slice(0, 200)}${e.content.length > 200 ? "..." : ""}`);
-                    return `${results.length} matches for "${args.query}":\n${lines.join("\n")}`;
-                },
-            }),
-            memory_read: tool({
-                description: "Read the full content of a specific memory file.",
-                args: {
-                    file_name: tool.schema.string().describe("File name of the memory to read (with or without .md extension)"),
-                },
-                async execute(args, _ctx) {
-                    const entry = readMemory(memoryRoot, args.file_name);
-                    if (!entry) {
-                        return `Memory "${args.file_name}" not found.`;
-                    }
-                    return `# ${entry.name}\n**Type:** ${entry.type}\n**Description:** ${entry.description}\n\n${entry.content}`;
-                },
-            }),
+        tool: buildMemoryTools(store, extraction),
+        dispose: async () => {
+            extraction.dispose();
+            owned.dispose();
         },
     };
 };
+export const MemoryPlugin = createMemoryPlugin();
+const plugin = { id: PLUGIN_ID, server: MemoryPlugin };
+export default plugin;
+export { MEMORY_AGENTS, MemoryOptionsSchema } from "./config.js";
+export { MemoryStore } from "./store/MemoryStore.js";