@yandy0725/pi-memory 1.3.1 → 1.3.3

package/index.ts

@@ -17,6 +17,15 @@ import { readDreamMeta, shouldNudge, writeDreamMeta } from "./src/nudge";
 import { resolveMemoryDir } from "./src/paths";
 import { searchSessions } from "./src/session-search";
 
+function extractAgentsMdBlocks(systemPrompt: string): string[] {
+	const blocks: string[] = [];
+	const re = /<project_instructions\s+path="([^"]+)">\n([\s\S]*?)<\/project_instructions>/g;
+	for (const match of systemPrompt.matchAll(re)) {
+		blocks.push(match[0]);
+	}
+	return blocks;
+}
+
 function resolveDefault(cfg: MemoryConfig, task: "dream" | "autoSurfacing" | "extractMemories", key: "model"): string | undefined;
 function resolveDefault(cfg: MemoryConfig, task: "dream" | "autoSurfacing" | "extractMemories", key: "sessionPersistence"): SessionPersistenceConfig | undefined;
 function resolveDefault(cfg: MemoryConfig, task: "dream" | "autoSurfacing" | "extractMemories", key: "model" | "sessionPersistence"): string | SessionPersistenceConfig | undefined {
@@ -31,6 +40,7 @@ export default function (pi: ExtensionAPI) {
 	let indexSnapshot = "";
 	let toolRegistered = false;
 	const injectedTopics = new Set<string>();
+	let lastSystemPrompt = "";
 
 	pi.on("session_start", async (_event, ctx) => {
 		config = await loadConfig(ctx);
@@ -91,6 +101,7 @@ export default function (pi: ExtensionAPI) {
 	});
 
 	pi.on("before_agent_start", async (event, ctx) => {
+		lastSystemPrompt = event.systemPrompt;
 		if (!config?.enabled || !indexSnapshot || !memoryDir) return;
 
 		const autoSurfacing = config.autoSurfacing;
@@ -151,6 +162,7 @@ export default function (pi: ExtensionAPI) {
 		if (!extractConfig?.enabled) return;
 		if (!event.messages || event.messages.length === 0) return;
 		runExtract({
+			agentsMdBlocks: extractAgentsMdBlocks(lastSystemPrompt),
 			model: resolveDefault(config, "extractMemories", "model"),
 			thinkLevel: extractConfig.thinkLevel,
 			memoryDir,

package/package.json

@@ -3,7 +3,7 @@
 	"publishConfig": {
 		"access": "public"
 	},
-	"version": "1.3.1",
+	"version": "1.3.3",
 	"description": "File-system driven persistent memory layer for pi coding agent",
 	"license": "MIT",
 	"repository": {

package/src/dream.ts

@@ -3,7 +3,7 @@ import type { ModelRegistry } from "@earendil-works/pi-coding-agent";
 import { runHeadlessAgent } from "./agent-runner";
 import type { SessionPersistenceConfig, ThinkLevel } from "./config";
 
-/** Build dream consolidation task. (unchanged) */
+/** Build dream consolidation task. */
 export function buildDreamTask(memoryDir: string, maxLines: number): string {
 	return `You are a memory consolidation agent. Your job is to read all memory files
 and consolidate them into a clean, deduplicated memory store.
@@ -33,6 +33,8 @@ Phase 4 — Prune & Index:
   description: specific summary that helps LLM match queries (be specific!)
   type: one of user, feedback, project, reference
   updated: today's date
+- Ensure entry titles are self-contained and descriptive
+  (only titles appear in future sessions' index, not the entry body)
 - Generate a compact hook (~150 chars) for each topic summarizing its entries
 - Rebuild MEMORY.md with one line per topic file (max ${maxLines} lines):
   - [Name](file.md) — hook
@@ -47,7 +49,13 @@ CRITICAL for hooks and descriptions:
 - Good: "SSH port 2222 on staging; MySQL 30s timeout; Redis auth fix"
 - Each topic file's \`## Entry Title\` blocks contain the actual memory entries.
   The MEMORY.md line is just a pointer — only ONE line per topic file.
-- When done, output a concise summary of changes (merged N, removed N, moved N, updated N).`;
+
+IMPORTANT — Do not prune process rules:
+- "Always do X" / "Never do Y" rules and workflow discipline entries are
+  as valuable as technical facts. Do not delete them as "obsolete"
+  just because they look like meta-instructions.
+
+When done, output a concise summary of changes (merged N, removed N, moved N, updated N).`;
 }
 
 export interface RunDreamOpts {

package/src/extract.ts

@@ -13,6 +13,7 @@ export interface RunExtractOpts {
 	parentModel?: Model<any>;
 	sessionPersistence?: SessionPersistenceConfig;
 	customTools?: ToolDefinition[];
+	agentsMdBlocks?: string[];  // 新增
 }
 
 /** Build extraction task prompt using memory tools instead of raw file I/O. */
@@ -20,6 +21,7 @@ export function buildExtractTask(
 	memoryDir: string,
 	messages: Array<{ role: string; content: string }>,
 	maxTokens: number,
+	agentsMdBlocks: string[] = [],  // 新增
 ): string {
 	const fromUser = messages.find((m) => m.role === "user");
 	const fromAssistant = messages.findLast((m) => m.role === "assistant");
@@ -35,34 +37,43 @@ export function buildExtractTask(
 		"",
 		"Analyze the conversation snippet below. If you find valuable learnings, persist them using the memory tools.",
 		"",
-		"You have these tools available:",
-		"- 'ls' and 'read': list files and read topic files in the memory directory to check for existing topics",
-		"- memory_search: search across all memory files for relevant existing entries",
-		"- memory_add: persist a new memory entry to a topic file (creates the topic if new)",
+		"## Tools",
+		"- ls — list files in the memory directory",
+		"- read — read MEMORY.md and topic files to check for existing topics",
+		"- memory_search — full-text search across all memory files for related entries",
+		"- memory_add — persist a new memory entry to a topic file (creates the topic if new)",
 		"",
-		"Use 'ls' to list files and 'read' to inspect MEMORY.md and topic files. Use memory_search to find related entries. Use memory_add to write new memories. Do NOT use 'bash', 'write', 'edit', or any other tools.",
+		"Do NOT use 'bash', 'write', 'edit', or any other tools.",
 		"",
-		"Worth remembering:",
-		"- User preferences, coding style choices, tooling preferences",
-		"- Project conventions, architecture decisions, naming patterns",
-		"- Debugging insights, workarounds, gotchas discovered",
-		'- "Always do X" / "Never do Y" rules',
-		"- References to external systems or documentation",
+		"## Workflow",
+		"1. Use ls + read to survey existing topic files and MEMORY.md index.",
+		"2. Use memory_search to check for overlapping or related entries.",
+		"3. Use memory_add to write new memories — only if the information is novel and valuable.",
 		"",
-		"NOT worth remembering:",
+		"## What to Remember",
+		"- Process rules: \"Always do X\" / \"Never do Y\" directives, workflow discipline, reporting standards, self-check habits — treat these as seriously as technical facts",
+		"- User preferences: coding style, tool choices, naming conventions, workflow habits",
+		"- Project conventions: architecture decisions, file organization, tech stack choices",
+		"- Discoveries: debugging workarounds, gotchas, configuration quirks, undocumented behavior",
+		"- References: external docs, APIs, or systems the user treats as important",
+		"- AGENTS.md rules that were violated in this conversation — extract for memory-level reinforcement (refer to the AGENTS.md content below)",
+		"",
+		"## What to Skip",
 		"- One-time task instructions or ephemeral details",
 		"- Code snippets or file paths derivable from the project",
-		"- Information already captured in CLAUDE.md or AGENTS.md",
+		"- AGENTS.md rules that were followed without issue (refer to the AGENTS.md content below)",
 		"- Git history or recent changes",
+		"- Obvious or trivial observations",
 		"",
-		"When writing memories:",
-		"- Use 'ls' and 'read' first to check for existing topic files and MEMORY.md index",
-		"- Use memory_search to find overlapping or related memories before adding",
-		"- Use descriptive, self-contained entry titles (only index lines are injected into future sessions)",
-		'- Choose the appropriate type: user, feedback, project, reference (default "feedback")',
-		"- Be concise but complete",
-		"- If unsure, do NOT write anything",
+		"## Memory Entry Guidelines",
+		"- Entry titles must be self-contained and descriptive (only titles appear in future sessions' index)",
+		'- Choose the appropriate type: user, feedback, project, or reference (default "feedback")',
+		"- Be concise but complete — one clear point per entry",
+		"- When in doubt, skip it",
 		"",
+		...(agentsMdBlocks.length > 0
+			? ["## AGENTS.md Rules", ...agentsMdBlocks, ""]
+			: []),
 		"=== Conversation ===",
 		`User: ${truncatedUser}`,
 		`Assistant: ${truncatedAssistant}`,
@@ -72,7 +83,7 @@ export function buildExtractTask(
 /** Fire-and-forget memory extraction. Does not await the headless agent. */
 export async function runExtract(opts: RunExtractOpts): Promise<void> {
 	if (opts.messages.length === 0) return;
-	const task = buildExtractTask(opts.memoryDir, opts.messages, opts.maxContextTokens);
+	const task = buildExtractTask(opts.memoryDir, opts.messages, opts.maxContextTokens, opts.agentsMdBlocks ?? []);
 	// fire-and-forget: runner disposes internally via finally
 	runHeadlessAgent({
 		task,

package/src/inject.ts

@@ -60,25 +60,7 @@ export async function scanTopics(memoryDir: string): Promise<TopicManifest[]> {
 	return manifests.sort((a, b) => b.mtimeMs - a.mtimeMs);
 }
 
-function buildSurfacingPrompt(manifest: TopicManifest[], userPrompt: string): string {
-	const lines = manifest.map((t) => {
-		return `[${t.type}] ${t.filename} — ${t.description.slice(0, 80)}`;
-	});
 
-	return [
-		"You are a memory relevance selector. Below is a list of memory topic files and a user message.",
-		"Select up to N topic files that are relevant to the user's current query.",
-		"Response format: JSON with a 'selected_files' array of filenames.",
-		"",
-		"=== Topic Files ===",
-		...lines,
-		"",
-		"=== User Message ===",
-		userPrompt,
-		"",
-		'Return: {"selected_files": ["a.md", "b.md"]}',
-	].join("\n");
-}
 
 export async function injectSurfacedContent(
 	memoryDir: string,
@@ -107,18 +89,27 @@ export async function injectSurfacedContent(
 	return `<relevant_memories>\n${blocks.join("\n\n")}\n</relevant_memories>`;
 }
 
-/** Build the side-query task prompt. Extracted for testability. */
-export function buildSideQueryTask(prompt: string, maxFiles: number): string {
+/** Build the side-query task prompt. */
+export function buildSideQueryTask(
+	manifest: TopicManifest[],
+	userPrompt: string,
+	maxFiles: number,
+): string {
+	const lines = manifest.map((t) =>
+		`[${t.type}] ${t.filename} — ${t.description.slice(0, 80)}`,
+	);
+
 	return [
-		"Respond with ONLY a JSON object.",
+		`You are a memory relevance selector. Select up to ${maxFiles} topic files most relevant to the user query.`,
+		"If nothing matches, select none.",
 		"",
-		"Below is a list of memory topic files and a user query.",
-		`Select up to ${maxFiles} topic files MOST relevant to the user's current query.`,
-		'If nothing is relevant, return {"selected_files": []}.',
+		"=== Topic Files ===",
+		...lines,
 		"",
-		prompt,
+		"=== User Query ===",
+		userPrompt,
 		"",
-		'Respond with EXACTLY: {"selected_files": [...]}',
+		'Respond with ONLY a JSON object: {"selected_files": ["filename.md", ...]}',
 	].join("\n");
 }
 
@@ -151,8 +142,7 @@ export async function runSideQuery(
 ): Promise<string[]> {
 	const candidates = manifest.filter((t) => !injectedTopics.has(t.filename));
 	if (candidates.length === 0) return [];
-	const surfacingPrompt = buildSurfacingPrompt(candidates, userPrompt);
-	const task = buildSideQueryTask(surfacingPrompt, maxFiles);
+	const task = buildSideQueryTask(candidates, userPrompt, maxFiles);
 	try {
 		const result = await runHeadlessAgent({
 			task,