feat(agent-sdk): let agent_knowledge_recall request source chunks

hindsight-docs/docs-integrations/openclaw.md

@@ -153,7 +153,7 @@ Optional settings in `~/.openclaw/openclaw.json`:
 - `enableKnowledgeTools` - Register `agent_knowledge_*` tools for explicit agent-driven lookup, reflection, ingest, and knowledge-page management (default: `false`).
 - `debug` - Enable debug logging (default: `false`).
 
-When using `agent_knowledge_recall` manually, pass `max_tokens` to control how much memory text the recall response may contain. The tool has no `max_results` parameter — to cap the number of automatically injected memories, use `recallTopK` on auto-recall instead.
+When using `agent_knowledge_recall` manually, pass `max_tokens` to control how much memory text the recall response may contain. The tool has no `max_results` parameter — to cap the number of automatically injected memories, use `recallTopK` on auto-recall instead. When the answer needs verbatim wording or an exact number rather than an extracted fact, pass `include_chunks: true` to also get the raw source text those memories came from, and `max_chunk_tokens` to bound it (default `8192`).
 
 When using `agent_knowledge_reflect`, keep the default conservative settings unless you intentionally need a deeper synthesis: `budget` defaults to `low`, `max_tokens` defaults to `1024`, and `fact_types` defaults to `world`, `experience`, and `observation`. Reflect calls can be more expensive than recall because they retrieve memories and then call the configured Reflect LLM to generate an answer. For production banks, set a finite bank-level `reflect_source_facts_max_tokens` value (for example `4096` or `8192`) instead of leaving it unlimited, so ad-hoc reflection cannot pull an unbounded amount of source facts into the LLM context.
 

hindsight-tools/hindsight-agent-sdk/src/index.ts

@@ -99,6 +99,14 @@ function parseRecallMaxTokens(input: unknown): number {
   return n;
 }
 
+/** Undefined on absent/invalid input so the client's own chunk budget default applies. */
+function parseChunkMaxTokens(input: unknown): number | undefined {
+  if (input === undefined || input === null) return undefined;
+  const n = Math.floor(Number(input));
+  if (!Number.isFinite(n) || n < 1) return undefined;
+  return n;
+}
+
 // ── Factory ────────────────────────────────────────────
 
 /**
@@ -261,6 +269,17 @@ export function createKnowledgeTools(opts: CreateKnowledgeToolsOptions): Knowled
             items: { type: "string", enum: ["world", "experience", "observation"] },
             description: "Alias for fact_types.",
           },
+          include_chunks: {
+            type: "boolean",
+            description:
+              "Also return the raw source text the memories were extracted from (default false). " +
+              "Use only when the answer needs verbatim wording, an exact quote, or an exact number " +
+              "that an extracted fact may have rounded or paraphrased — it costs significantly more tokens.",
+          },
+          max_chunk_tokens: {
+            type: "number",
+            description: "Token budget for the returned source chunks (default 8192).",
+          },
         },
         required: ["query"],
       },
@@ -273,6 +292,8 @@ export function createKnowledgeTools(opts: CreateKnowledgeToolsOptions): Knowled
         const result = await client.recall(bankId, params.query as string, {
           maxTokens,
           types,
+          includeChunks: params.include_chunks === true,
+          maxChunkTokens: parseChunkMaxTokens(params.max_chunk_tokens),
         });
         return ok(result);
       },

hindsight-tools/hindsight-agent-sdk/tests/index.test.ts

@@ -206,6 +206,63 @@ describe("createKnowledgeTools", () => {
     expect(body.max_tokens).toBe(512);
   });
 
+  it("recall omits source chunks unless include_chunks is exactly true", async () => {
+    const tool = tools.find((t) => t.name === "agent_knowledge_recall")!;
+
+    // Omitted, false, and a truthy-string "false" must all leave chunks off.
+    for (const params of [
+      { query: "no chunks" },
+      { query: "no chunks", include_chunks: false },
+      { query: "no chunks", include_chunks: "false" },
+      // A chunk budget alone must not switch chunks on.
+      { query: "no chunks", max_chunk_tokens: 2048 },
+    ]) {
+      mockFetch.mockReset();
+      mockFetch.mockReturnValueOnce(mockResponse({ results: [{ text: "found" }] }));
+      await tool.execute(params);
+      const body = await getBody(mockFetch.mock.calls[0]);
+      expect(body.include?.chunks, JSON.stringify(params)).toBeUndefined();
+    }
+  });
+
+  it("recall include_chunks requests source chunks and returns them to the agent", async () => {
+    mockFetch.mockReturnValueOnce(
+      mockResponse({
+        results: [{ text: "found" }],
+        chunks: { c1: { chunk_text: "the fee is 4.375%", truncated: false } },
+      })
+    );
+
+    const tool = tools.find((t) => t.name === "agent_knowledge_recall")!;
+    const result = await tool.execute({ query: "exact wording", include_chunks: true });
+
+    const body = await getBody(mockFetch.mock.calls[0]);
+    expect(body.include.chunks).toEqual({ max_tokens: 8192 });
+    // The chunks must survive into the tool output, not just the request.
+    expect(JSON.parse(result.content[0].text).chunks.c1.chunk_text).toBe("the fee is 4.375%");
+  });
+
+  it("recall falls back to the client's chunk budget on unusable max_chunk_tokens", async () => {
+    const tool = tools.find((t) => t.name === "agent_knowledge_recall")!;
+
+    for (const [maxChunkTokens, expected] of [
+      [2048, 2048],
+      [0, 8192],
+      [-5, 8192],
+      ["not-a-number", 8192],
+    ] as const) {
+      mockFetch.mockReset();
+      mockFetch.mockReturnValueOnce(mockResponse({ results: [{ text: "found" }] }));
+      await tool.execute({
+        query: "exact wording",
+        include_chunks: true,
+        max_chunk_tokens: maxChunkTokens,
+      });
+      const body = await getBody(mockFetch.mock.calls[0]);
+      expect(body.include.chunks.max_tokens, String(maxChunkTokens)).toBe(expected);
+    }
+  });
+
   it("reflect sends POST with conservative defaults", async () => {
     mockFetch.mockReturnValueOnce(mockResponse({ text: "answer" }));
 

skills/hindsight-docs/references/sdks/integrations/openclaw.md

@@ -148,7 +148,7 @@ Optional settings in `~/.openclaw/openclaw.json`:
 - `enableKnowledgeTools` - Register `agent_knowledge_*` tools for explicit agent-driven lookup, reflection, ingest, and knowledge-page management (default: `false`).
 - `debug` - Enable debug logging (default: `false`).
 
-When using `agent_knowledge_recall` manually, pass `max_tokens` to control how much memory text the recall response may contain. The tool has no `max_results` parameter — to cap the number of automatically injected memories, use `recallTopK` on auto-recall instead.
+When using `agent_knowledge_recall` manually, pass `max_tokens` to control how much memory text the recall response may contain. The tool has no `max_results` parameter — to cap the number of automatically injected memories, use `recallTopK` on auto-recall instead. When the answer needs verbatim wording or an exact number rather than an extracted fact, pass `include_chunks: true` to also get the raw source text those memories came from, and `max_chunk_tokens` to bound it (default `8192`).
 
 When using `agent_knowledge_reflect`, keep the default conservative settings unless you intentionally need a deeper synthesis: `budget` defaults to `low`, `max_tokens` defaults to `1024`, and `fact_types` defaults to `world`, `experience`, and `observation`. Reflect calls can be more expensive than recall because they retrieve memories and then call the configured Reflect LLM to generate an answer. For production banks, set a finite bank-level `reflect_source_facts_max_tokens` value (for example `4096` or `8192`) instead of leaving it unlimited, so ad-hoc reflection cannot pull an unbounded amount of source facts into the LLM context.