airkjw/claude-mem
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix(openclaw): inject context via system prompt instead of overwriting MEMORY.md (#1386)

Browse Source 
* fix(openclaw): inject context via system prompt instead of overwriting MEMORY.md

The OpenClaw plugin was overwriting each agent's MEMORY.md with a large
auto-generated observation dump (~12-15KB) on every before_agent_start
and tool_result_persist event. This conflicts with OpenClaw's design
where MEMORY.md is agent-curated long-term memory.

Migrate context injection from file-based (writeFile MEMORY.md) to
OpenClaw's native before_prompt_build hook, which returns context via
appendSystemContext. This keeps MEMORY.md under agent control while
still providing cross-session observation context to the LLM.

Changes:
- Add before_prompt_build hook that returns { appendSystemContext }
- Remove writeFile/MEMORY.md sync from before_agent_start
- Remove MEMORY.md sync from tool_result_persist (observations still recorded)
- Add 60s TTL cache to avoid re-fetching context on every LLM turn
- Add syncMemoryFileExclude config for per-agent opt-out
- Remove dead workspaceDirsBySessionKey tracking map
- Rewrite test suite to verify prompt injection instead of file writes

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix(ui): align settings defaults with backend and use nullish coalescing

The web UI had two issues causing settings inflation:

1. DEFAULT_SETTINGS in the UI used FULL_COUNT='5' and all token columns
   'true', while SettingsDefaultsManager (backend) uses FULL_COUNT='0'
   and token columns 'false'. Opening the settings modal and saving
   without changes would silently inflate the context.

2. useSettings used || for fallback, which treats '0' and 'false' as
   falsy — even when the backend correctly returns these values, the UI
   would replace them with inflated defaults. Changed to ?? (nullish
   coalescing) so only null/undefined trigger the fallback.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs(openclaw): update integration docs for system prompt injection

Reflect the migration from MEMORY.md file writes to before_prompt_build
hook-based context injection:

- Update architecture diagram and overview to show new hook flow
- Replace "MEMORY.md Live Sync" section with "System Prompt Context Injection"
- Update event lifecycle steps (before_agent_start, tool_result_persist)
- Add before_prompt_build step with TTL cache description
- Document new syncMemoryFileExclude config parameter
- Update session tracking to reflect removed workspaceDirsBySessionKey

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs: fix terminology and update SKILL.md for system prompt injection

Replace "prompt injection" with "context injection" in docs to avoid
confusion with the OWASP security term. Update openclaw/SKILL.md to
reflect the new before_prompt_build hook and remove stale MEMORY.md
references.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-authored-by: Alex Newman <thedotmack@gmail.com>
This commit is contained in:
Glucksberg
2026-03-17 20:14:30 -04:00
committed by GitHub
parent 9e7b08445f
commit 9361e33b6d
7 changed files with 267 additions and 204 deletions
Download Patch File Download Diff File Expand all files Collapse all files
openclaw/src/index.ts
+66 -31
View File
@@ -1,5 +1,5 @@
import { writeFile } from "fs/promises";
import { join } from "path";
// No file-system imports needed — context is injected via system prompt hook,
// not by writing to MEMORY.md.
// Minimal type declarations for the OpenClaw Plugin SDK.
// These match the real OpenClawPluginApi provided by the gateway at runtime.
@@ -35,6 +35,18 @@ interface BeforeAgentStartEvent {
prompt?: string;
}
interface BeforePromptBuildEvent {
prompt: string;
messages: unknown[];
}
interface BeforePromptBuildResult {
systemPrompt?: string;
prependContext?: string;
prependSystemContext?: string;
appendSystemContext?: string;
}
interface ToolResultPersistEvent {
toolName?: string;
params?: Record<string, unknown>;
@@ -87,6 +99,7 @@ interface MessageContext {
}
type EventCallback<T> = (event: T, ctx: EventContext) => void | Promise<void>;
type PromptBuildCallback = (event: BeforePromptBuildEvent, ctx: EventContext) => BeforePromptBuildResult | Promise<BeforePromptBuildResult | void> | void;
type MessageEventCallback<T> = (event: T, ctx: MessageContext) => void | Promise<void>;
interface OpenClawPluginApi {
@@ -109,7 +122,8 @@ interface OpenClawPluginApi {
requireAuth?: boolean;
handler: (ctx: PluginCommandContext) => PluginCommandResult | Promise<PluginCommandResult>;
}) => void;
on: ((event: "before_agent_start", callback: EventCallback<BeforeAgentStartEvent>) => void) &
on: ((event: "before_prompt_build", callback: PromptBuildCallback) => void) &
((event: "before_agent_start", callback: EventCallback<BeforeAgentStartEvent>) => void) &
((event: "tool_result_persist", callback: EventCallback<ToolResultPersistEvent>) => void) &
((event: "agent_end", callback: EventCallback<AgentEndEvent>) => void) &
((event: "session_start", callback: EventCallback<SessionStartEvent>) => void) &
@@ -166,6 +180,7 @@ interface FeedEmojiConfig {
interface ClaudeMemPluginConfig {
syncMemoryFile?: boolean;
syncMemoryFileExclude?: string[];
project?: string;
workerPort?: number;
observationFeed?: {
@@ -532,8 +547,8 @@ export default function claudeMemPlugin(api: OpenClawPluginApi): void {
// Session tracking for observation I/O
// ------------------------------------------------------------------
const sessionIds = new Map<string, string>();
const workspaceDirsBySessionKey = new Map<string, string>();
const syncMemoryFile = userConfig.syncMemoryFile !== false; // default true
const syncMemoryFileExclude = new Set(userConfig.syncMemoryFileExclude || []);
function getContentSessionId(sessionKey?: string): string {
const key = sessionKey || "default";
@@ -543,27 +558,45 @@ export default function claudeMemPlugin(api: OpenClawPluginApi): void {
return sessionIds.get(key)!;
}
async function syncMemoryToWorkspace(workspaceDir: string, ctx?: EventContext): Promise<void> {
function shouldInjectContext(ctx?: EventContext): boolean {
if (!syncMemoryFile) return false;
const agentId = ctx?.agentId;
if (agentId && syncMemoryFileExclude.has(agentId)) return false;
return true;
}
// TTL cache for context injection to avoid re-fetching on every LLM turn.
// before_prompt_build fires on every turn; caching for 60s keeps the worker
// load manageable while still picking up new observations reasonably quickly.
const CONTEXT_CACHE_TTL_MS = 60_000;
const contextCache = new Map<string, { text: string; fetchedAt: number }>();
async function getContextForPrompt(ctx?: EventContext): Promise<string | null> {
// Include both the base project and agent-scoped project (e.g. "openclaw" + "openclaw-main")
const projects = [baseProjectName];
const agentProject = ctx ? getProjectName(ctx) : null;
if (agentProject && agentProject !== baseProjectName) {
projects.push(agentProject);
}
const cacheKey = projects.join(",");
// Return cached context if still fresh
const cached = contextCache.get(cacheKey);
if (cached && Date.now() - cached.fetchedAt < CONTEXT_CACHE_TTL_MS) {
return cached.text;
}
const contextText = await workerGetText(
workerPort,
`/api/context/inject?projects=${encodeURIComponent(projects.join(","))}`,
`/api/context/inject?projects=${encodeURIComponent(cacheKey)}`,
api.logger
);
if (contextText && contextText.trim().length > 0) {
try {
await writeFile(join(workspaceDir, "MEMORY.md"), contextText, "utf-8");
api.logger.info(`[claude-mem] MEMORY.md synced to ${workspaceDir}`);
} catch (writeError: unknown) {
const msg = writeError instanceof Error ? writeError.message : String(writeError);
api.logger.warn(`[claude-mem] Failed to write MEMORY.md: ${msg}`);
}
const trimmed = contextText.trim();
contextCache.set(cacheKey, { text: trimmed, fetchedAt: Date.now() });
return trimmed;
}
return null;
}
// ------------------------------------------------------------------
@@ -611,14 +644,9 @@ export default function claudeMemPlugin(api: OpenClawPluginApi): void {
});
// ------------------------------------------------------------------
// Event: before_agent_start — init session + sync MEMORY.md + track workspace
// Event: before_agent_start — init session
// ------------------------------------------------------------------
api.on("before_agent_start", async (event, ctx) => {
// Track workspace dir so tool_result_persist can sync MEMORY.md later
if (ctx.workspaceDir) {
workspaceDirsBySessionKey.set(ctx.sessionKey || "default", ctx.workspaceDir);
}
// Initialize session in the worker so observations are not skipped
// (the privacy check requires a stored user prompt to exist)
const contentSessionId = getContentSessionId(ctx.sessionKey);
@@ -627,15 +655,28 @@ export default function claudeMemPlugin(api: OpenClawPluginApi): void {
project: getProjectName(ctx),
prompt: event.prompt || "agent run",
}, api.logger);
});
// Sync MEMORY.md before agent runs (provides context to agent)
if (syncMemoryFile && ctx.workspaceDir) {
await syncMemoryToWorkspace(ctx.workspaceDir, ctx);
// ------------------------------------------------------------------
// Event: before_prompt_build — inject context into system prompt
//
// Instead of writing to MEMORY.md (which conflicts with agent-curated
// memory), inject the observation timeline via appendSystemContext.
// This keeps MEMORY.md under the agent's control while still providing
// cross-session context to the LLM.
// ------------------------------------------------------------------
api.on("before_prompt_build", async (_event, ctx) => {
if (!shouldInjectContext(ctx)) return;
const contextText = await getContextForPrompt(ctx);
if (contextText) {
api.logger.info(`[claude-mem] Context injected via system prompt for agent=${ctx.agentId ?? "unknown"}`);
return { appendSystemContext: contextText };
}
});
// ------------------------------------------------------------------
// Event: tool_result_persist — record tool observations + sync MEMORY.md
// Event: tool_result_persist — record tool observations
// ------------------------------------------------------------------
api.on("tool_result_persist", (event, ctx) => {
api.logger.info(`[claude-mem] tool_result_persist fired: tool=${event.toolName ?? "unknown"} agent=${ctx.agentId ?? "none"} session=${ctx.sessionKey ?? "none"}`);
@@ -663,7 +704,7 @@ export default function claudeMemPlugin(api: OpenClawPluginApi): void {
toolResponseText = toolResponseText.slice(0, MAX_TOOL_RESPONSE_LENGTH);
}
// Fire-and-forget: send observation + sync MEMORY.md in parallel
// Fire-and-forget: send observation to worker
workerPostFireAndForget(workerPort, "/api/sessions/observations", {
contentSessionId,
tool_name: toolName,
@@ -671,11 +712,6 @@ export default function claudeMemPlugin(api: OpenClawPluginApi): void {
tool_response: toolResponseText,
cwd: "",
}, api.logger);
const workspaceDir = ctx.workspaceDir || workspaceDirsBySessionKey.get(ctx.sessionKey || "default");
if (syncMemoryFile && workspaceDir) {
syncMemoryToWorkspace(workspaceDir, ctx);
}
});
// ------------------------------------------------------------------
@@ -722,15 +758,14 @@ export default function claudeMemPlugin(api: OpenClawPluginApi): void {
api.on("session_end", async (_event, ctx) => {
const key = ctx.sessionKey || "default";
sessionIds.delete(key);
workspaceDirsBySessionKey.delete(key);
});
// ------------------------------------------------------------------
// Event: gateway_start — clear session tracking for fresh start
// ------------------------------------------------------------------
api.on("gateway_start", async () => {
workspaceDirsBySessionKey.clear();
sessionIds.clear();
contextCache.clear();
api.logger.info("[claude-mem] Gateway started — session tracking reset");
});