Skip to content

long-term-memory.md

Latest commit

 

History

History
202 lines (140 loc) · 6.94 KB

long-term-memory.md

File metadata and controls

202 lines (140 loc) · 6.94 KB
description Taskade's Long-Term Memory architecture — memory that lives as real Projects in your workspace, editable, browsable, and API-addressable.

Long-Term Memory

Taskade EVE's long-term memory is a Project in your workspace. Not an opaque blob. You can browse it, edit it, version it, and query it via the Public API — just like any other project.

This architectural choice is what makes Taskade memory transparent, controllable, and composable with the rest of the platform.

Table of Contents

Mental Model — Memory as Projects

Most AI platforms store memory in a hidden vector database. You can't see what the agent remembers or correct it directly.

Taskade takes a different approach.

graph TB
    subgraph "Traditional AI Platforms"
        A1[Agent] -.->|opaque blob| M1[(Hidden<br/>memory)]
    end

    subgraph "Taskade — Memory as Projects"
        A2[Taskade EVE]
        M2[Memory Project<br/>in your workspace]
        U2[User]

        A2 -.writes.-> M2
        A2 -.reads.-> M2
        U2 -.can browse.-> M2
        U2 -.can edit.-> M2
        U2 -.can query via API.-> M2
    end
Loading

Why this matters:

  • Transparency — you see exactly what EVE remembers.
  • Controllability — edit or delete memory blocks like any project content.
  • Composability — use memory as an input to automations, agents, or other apps.
  • Portability — memory is part of your workspace, not locked in a separate system.

How EVE Uses Memory

EVE is Taskade's meta-agent — the one you talk to when building apps and orchestrating work.

  • Memory persists across sessions. EVE remembers what you've been working on even after you close and reopen Taskade.
  • Memory is workspace-scoped. Each workspace has its own EVE memory.
  • Memory grows automatically as you interact — no manual tagging required.
  • EVE runs on the same Workspace DNA (Projects + Agents + Automations) it orchestrates for you.

Agent Memory vs EVE Memory

Taskade has two distinct memory systems. Both use the same architectural principle (memory-as-Projects), but they serve different purposes.

Dimension Per-Agent Memory EVE Workspace Memory
Scope Single agent Whole workspace
Persists across Conversations with that agent All sessions and agents
User-editable Limited Full — it's a Project
API addressable Via agent endpoints Via project endpoints
Model support All frontier models All frontier models
Typical use Personalized assistant behavior Cross-session workspace context

Model Agnosticism

Memory works across every available AI model.

  • Persistent memory is not limited to specific models — every AI agent supports memory regardless of which model you pick.
  • The memory tool is enabled for all 11+ frontier models currently integrated with Taskade.
  • Switching model (say, from Claude to Gemini) does not reset an agent's memory.

Inspecting and Editing Memory

Finding the memory project

In your workspace sidebar, look for the project prefixed with "EVE Memory" or your agent's memory project (naming varies based on when it was created).

Editing manually

Open the memory project like any other project. You can:

  • Remove entries you don't want remembered.
  • Correct facts EVE got wrong.
  • Add context manually when you want EVE to know something it hasn't inferred yet.
  • Version the memory (Taskade's version history applies).

{% hint style="info" %} EVE reads from the memory project at each interaction, so your edits take effect immediately in the next conversation. {% endhint %}

Common cleanup patterns

  • After a major project pivot, prune outdated assumptions.
  • Before sharing a workspace, review EVE memory for anything sensitive.
  • Periodically archive resolved context so memory stays focused on current work.

API Access

Because memory is a Project, it's addressable via the Public API.

Find the memory project

import { Taskade } from "@taskade/sdk";

const taskade = new Taskade({ token: process.env.TASKADE_TOKEN! });

const projects = await taskade.projects.list({ workspaceId: WORKSPACE_ID });
const memory = projects.items.find(p => p.name.startsWith("EVE Memory"));

Read memory contents

if (memory) {
  const content = await taskade.projects.get(memory.id);
  console.log("EVE remembers:", content.blocks);
}

Attach memory as knowledge to another agent

// Give a specialist agent access to workspace context
await taskade.agents.addProjectKnowledge(AGENT_ID, {
  projectId: memory.id,
});

{% hint style="info" %} Memory projects follow the same schema as any other project: blocks, tasks, metadata. Treat them like a project you happen to read a lot. {% endhint %}

Privacy and Scoping

  • Workspace-scoped. EVE memory does not leak across workspaces.
  • Never in exported bundles. When you export an app via GitHub or Bundles, memory is not included.
  • Controlled by workspace members. Anyone with edit access to the workspace can view and modify memory.
  • No external service. Memory lives in Taskade's infrastructure, not a third-party vector store.

Best Practices

  • Let EVE remember enough to be useful. Resist the instinct to wipe memory frequently — context compounds.
  • Prune when context drifts. Major pivots are a good time to review and edit.
  • Use memory as input to automations. Read memory via the API to personalize triggered flows.
  • Scope sensitive info. Keep confidential context in a dedicated workspace if you're collaborating with external parties.
  • Document memory assumptions in the project itself. EVE reads it — so add explicit "please remember that…" notes as project content.

Related

{% content-ref url="api-v2-reference.md" %} api-v2-reference.md {% endcontent-ref %}

{% content-ref url="sdk-cookbook.md" %} sdk-cookbook.md {% endcontent-ref %}

{% content-ref url="autonomous-agents.md" %} autonomous-agents.md {% endcontent-ref %}

{% content-ref url="../genesis-living-system-builder/ai-features/agent-knowledge.md" %} agent-knowledge.md {% endcontent-ref %}