Add memory agent and caching functionality 🧠✨

docs/src/components/BuiltinAgents.mdx

@@ -11,4 +11,5 @@ import { LinkCard } from '@astrojs/starlight/components';
 <LinkCard title="agent git" description="query a repository using Git to accomplish tasks. Provide all the context information available to execute git queries." href="/genaiscript/reference/scripts/system#systemagent_git" />
 <LinkCard title="agent github" description="query GitHub to accomplish tasks" href="/genaiscript/reference/scripts/system#systemagent_github" />
 <LinkCard title="agent interpreter" description="run code interpreters for Python, Math. Use this agent to ground computation questions." href="/genaiscript/reference/scripts/system#systemagent_interpreter" />
+<LinkCard title="agent memory" description="queries the memories created by other agent conversations." href="/genaiscript/reference/scripts/system#systemagent_memory" />
 <LinkCard title="agent user_input" description="ask user for input to confirm, select or answer the question in the query. The message should be very clear and provide all the context." href="/genaiscript/reference/scripts/system#systemagent_user_input" />

docs/src/content/docs/reference/scripts/agents.mdx

@@ -148,3 +148,15 @@
 ```
 
 This full source of this agent is defined in the [system.agent_github](/genaiscript/reference/scripts/system/#systemagent_github) system prompt.
+
+## Agent Memory
+
+The `agent_memory` is a special agent that queries the memories created by other agent conversations. It is used to store and retrieve information from the LLM's memory.
+
+All agent contribute to the conversation memory, and use the `agent_memory`, tool unless it is explicitely disabled using `disableMemory`.
+
+```js "disableMemory: true"
+defAgent(..., { disableMemory: true })
+```
+
+To enable agent memory in the top level script, add the `agent_memory` tool.

docs/src/content/docs/reference/scripts/system.mdx

@@ -234,8 +234,7 @@
     "github",
     "query GitHub to accomplish tasks",
     `Your are a helpfull LLM agent that can query GitHub to accomplish tasks. Answer the question in QUERY.
     - Prefer diffing job logs rather downloading entire logs which can be very large.
-    - Pull Requests ar a specialized type of issues.
     `,
     {
         model,
@@ -290,6 +289,57 @@
 `````
 
 
+### `system.agent_memory`
+
+agent that retrieves memories
+
+
+
+
+
+`````js wrap title="system.agent_memory"
+system({
+    title: "agent that retrieves memories",
+})
+
+const cache = await host.cache("memory")
+defAgent(
+    "memory",
+    "queries the memories created by other agent conversations.",
+    async (ctx) => {
+        const memories = await cache.values()
+        ctx.$`Your are a helpfull LLM agent that acts as a knowledge base for memories created by other agents.
+
+    Answer the question in QUERY with the memories in MEMORY.
+
+    - Use the information in MEMORY exclusively to answer the question in QUERY.
+    - If the information in MEMORY is not enough to answer the question in QUERY, respond <NO_MEMORY>.
+    - The memory 
+    `
+        memories.reverse().forEach(
+            ({ agent, query, answer }) =>
+                ctx.def(
+                    "MEMORY",
+                    `${agent}> ${query}?
+                ${answer}
+                `
+                ),
+            {
+                flex: 1,
+            }
+        )
+    },
+    {
+        model: "openai:gpt-4o",
+        flexTokens: 30000,
+        system: ["system"],
+        disableMemory: true,
+    }
+)
+
+`````
+
+
 ### `system.agent_user_input`
 
 Agent that can asks questions to the user.

packages/core/src/constants.ts

@@ -255,3 +255,5 @@ export const CLI_ENV_VAR_RX = /^genaiscript_var_/i
 
 export const GIT_DIFF_MAX_TOKENS = 8000
 export const MAX_TOOL_CONTENT_TOKENS = 4000
+
+export const MEMORY_CACHE_NAME = "memory"

packages/core/src/genaisrc/system.agent_github.genai.mjs

@@ -9,7 +9,6 @@ defAgent(
     "query GitHub to accomplish tasks",
     `Your are a helpfull LLM agent that can query GitHub to accomplish tasks. Answer the question in QUERY.
     - Prefer diffing job logs rather downloading entire logs which can be very large.
-    - Pull Requests ar a specialized type of issues.
     `,
     {
         model,

packages/core/src/genaisrc/system.agent_memory.genai.mjs

@@ -0,0 +1,38 @@
+system({
+    title: "agent that retrieves memories",
+})
+
+const cache = await host.cache("memory")
+defAgent(
+    "memory",
+    "queries the memories created by other agent conversations.",
+    async (ctx) => {
+        const memories = await cache.values()
+        ctx.$`Your are a helpfull LLM agent that acts as a knowledge base for memories created by other agents.
+
+    Answer the question in QUERY with the memories in MEMORY.
+
+    - Use the information in MEMORY exclusively to answer the question in QUERY.
+    - If the information in MEMORY is not enough to answer the question in QUERY, respond <NO_MEMORY>.
+    - The memory 
+    `
+        memories.reverse().forEach(
+            ({ agent, query, answer }) =>
+                ctx.def(
+                    "MEMORY",
+                    `${agent}> ${query}?
+                ${answer}
+                `
+                ),
+            {
+                flex: 1,
+            }
+        )
+    },
+    {
+        model: "openai:gpt-4o",
+        flexTokens: 30000,
+        system: ["system"],
+        disableMemory: true,
+    }
+)

packages/core/src/runpromptcontext.ts

@@ -41,6 +41,7 @@
 import { parseModelIdentifier, resolveModelConnectionInfo } from "./models"
 import {
     CHAT_REQUEST_PER_MODEL_CONCURRENT_LIMIT,
+    MEMORY_CACHE_NAME,
     MODEL_PROVIDER_AICI,
     SYSTEM_FENCE,
 } from "./constants"
@@ -59,6 +60,7 @@
 import { dedent } from "./indent"
 import { runtimeHost } from "./host"
 import { writeFileEdits } from "./fileedits"
+import { MemoryCache } from "./cache"
 
 export function createChatTurnGenerationContext(
     options: GenerationOptions,
@@ -313,7 +315,8 @@
         ) => Promise<void>,
         options?: DefAgentOptions
     ): void => {
-        const { tools, system, ...rest } = options || {}
+        const { tools, system, disableMemory, ...rest } = options || {}
+        const memory = !disableMemory
 
         name = name.replace(/^agent_/i, "")
         const agentName = `agent_${name}`
@@ -327,7 +330,9 @@
         const agentTools = resolveTools(
             runtimeHost.project,
             agentSystem,
-            arrayify(tools)
+            [...arrayify(tools), memory ? "agent_memory" : undefined].filter(
+                (t) => !!t
+            )
         )
         const agentDescription = dedent`Agent that uses an LLM to ${description}.\nAvailable tools: 
         ${agentTools.map((t) => `- ${t.description}`).join("\n")}` // DO NOT LEAK TOOL ID HERE
@@ -347,17 +352,40 @@
             },
             async (args) => {
                 const { context, query } = args
-                logVerbose(`${agentLabel}: ${query}`)
+                context.log(`${agentLabel}: ${query}`)
                 const res = await runPrompt(
                     async (_) => {
                         _.def("QUERY", query)
                         if (typeof fn === "string") _.writeText(dedent(fn))
                         else await fn(_, args)
                         _.$`- Assume that your answer will be analyzed by an LLM, not a human.
+                ${memory ? `- If you are missing information, try querying the memory using 'agent_memory'.` : ""}                        
                 - If you are missing information, reply "MISSING_INFO: <what is missing>".
                 - If you cannot answer the query, return "NO_ANSWER: <reason>".
                 - Be concise. Minimize output to the most relevant information to save context tokens.
                 `
+                        if (memory)
+                            _.defOutputProcessor(async ({ text }) => {
+                                const agentMemory = MemoryCache.byName<
+                                    { agent: string; query: string },
+                                    {
+                                        agent: string
+                                        query: string
+                                        answer: string
+                                    }
+                                >(MEMORY_CACHE_NAME)
+                                const cacheKey = { agent: agentName, query }
+                                const cachedValue = {
+                                    ...cacheKey,
+                                    answer: text,
+                                }
+                                await agentMemory.set(cacheKey, cachedValue)
+                                trace.detailsFenced(
+                                    `🧠 memory: ${query}`,
+                                    cachedValue.answer,
+                                    "markdown"
+                                )
+                            })
                     },
                     {
                         label: agentLabel,

packages/core/src/systems.ts

@@ -42,7 +42,7 @@ export function resolveSystems(
         if (/\W(diagram|chart)\W/i.test(jsSource))
             systems.push("system.diagrams")
 
-        if (/github/i.test(jsSource)) systems.push("system.github_info")
+        if (/\W(github)\W/i.test(jsSource)) systems.push("system.github_info")
     }
 
     // Include tools-related systems if specified in the script

packages/core/src/types/prompt_template.d.ts

@@ -2074,7 +2074,12 @@ interface DefToolOptions {
     maxTokens?: number
 }
 
-interface DefAgentOptions extends Omit<PromptGeneratorOptions, "label"> {}
+interface DefAgentOptions extends Omit<PromptGeneratorOptions, "label"> {
+    /**
+     * Excludes agent conversation from agent memory
+     */
+    disableMemory?: boolean
+}
 
 type ChatAgentHandler = (
     ctx: ChatGenerationContext,