memory_save

Instructions

Implementation Reference

• packages/server/src/tools/memory-save.ts:14-33 (handler)

  The handler function that executes the memory_save tool logic. It generates an ID, embeds the content via Ollama, inserts a row into the SQLite memories table, and returns a confirmation string.

  export async function handleMemorySave(
    content: string,
    category: MemoryCategory = "general",
    project?: string,
  ): Promise<string> {
    const db = getDb();
    const now = Date.now();
    const id = generateId("mem");
  
    const embedding = await embed(content);
    const embeddingBuf = embedding ? embeddingToBuffer(embedding) : null;
  
    db.prepare(
      `INSERT INTO memories (id, content, category, project, created_at, updated_at, accessed_at, embedding)
       VALUES (?, ?, ?, ?, ?, ?, ?, ?)`,
    ).run(id, content, category, project ?? null, now, now, now, embeddingBuf);
  
    const projectStr = project ? ` [${project}]` : "";
    return `Saved memory ${id}${projectStr} (${category}):\n${content.slice(0, 100)}${content.length > 100 ? "..." : ""}`;
  }

• packages/server/src/tools/memory-save.ts:4-12 (schema)

  Type definition for MemoryCategory, used as input schema for the handler function.

  export type MemoryCategory =
    | "decision"
    | "preference"
    | "fact"
    | "episode"
    | "lesson"
    | "architecture"
    | "framework"
    | "general";

• packages/server/src/index.ts:247-269 (registration)

  Registration of the memory_save tool with the MCP server, including Zod schema for content, category, and project inputs.

  server.tool(
    "memory_save",
    "Save a fact, decision, preference, or lesson to long-term memory. Automatically generates an embedding for future semantic search.",
    {
      content: z.string().describe("The memory content — be specific and self-contained"),
      category: z
        .enum(["decision", "preference", "fact", "episode", "lesson", "architecture", "framework", "general"])
        .optional()
        .describe("Memory category (default: general)"),
      project: z.string().optional().describe("Associated project name"),
    },
    async ({ content, category, project }) => {
      try {
        const result = await handleMemorySave(content, category, project);
        return { content: [{ type: "text", text: result }] };
      } catch (err) {
        return {
          content: [{ type: "text", text: `Error saving memory: ${err}` }],
          isError: true,
        };
      }
    },
  );

• packages/server/src/memory/db.ts:84-88 (helper)

  Helper function that generates unique IDs with a prefix (used as 'mem-...' for memory entries).

  export function generateId(prefix: string): string {
    const ts = Date.now();
    const rand = Math.random().toString(36).slice(2, 6);
    return `${prefix}-${ts}-${rand}`;
  }

• packages/server/src/memory/embeddings.ts:53-60 (helper)

  Helper that generates an embedding vector for memory content using Ollama's nomic-embed-text model.

  export async function embed(text: string): Promise<Float32Array | null> {
    const available = await ensureOllama();
    if (!available) return null;
  
    const client = getClient();
    const response = await client.embed({ model: MODEL, input: text });
    return new Float32Array(response.embeddings[0]);
  }