Additional

Supermemory Integration

Synced from github.com/CoWork-OS/CoWork-OS/docs

CoWork OS can use Supermemory as an external memory layer alongside its built-in local memory runtime.

This integration is intentionally modeled after the Hermes-style provider shape:

native provider-style configuration in Settings → Memory Hub
workspace-scoped container tags
explicit external memory tools
optional prompt-time profile injection
optional background mirroring of local memory writes
guarded failure behavior so provider outages do not break the main agent loop

Supermemory does not replace CoWork's local memory system. CoWork keeps its own archive memory, curated hot memory, workspace kit files, transcript recall, and knowledge graph. Supermemory is an additional external memory lane.

What It Adds

When enabled, CoWork adds four explicit tools to the agent runtime:

supermemory_profile
supermemory_search
supermemory_remember
supermemory_forget

It also adds two optional runtime behaviors:

Prompt profile injection: fetches a scoped Supermemory profile and appends it as soft context during chat, execution, and follow-up turns
Memory mirroring: mirrors non-private CoWork memory captures into Supermemory as indexed external documents

Setup

Open Settings → Memory Hub.
Find the Supermemory section.
Enable Supermemory.
Paste your Supermemory API key.
Leave the default base URL unless you are self-hosting.
Choose a container-tag template.
Save settings.
Click Test Connection.

The default base URL is:

https://api.supermemory.ai

The default container template is:

cowork:{workspaceId}

Supported template variables:

{workspaceId}
{workspaceName}

These are sanitized into a valid Supermemory containerTag.

Runtime Model

CoWork now has three distinct memory surfaces:

Local prompt-visible memory Curated hot memory, USER.md, MEMORY.md, and the L0/L1 wake-up layers.
Local deep recall search_memories, search_sessions, search_quotes, memory_topics_load, transcript spans, and workspace markdown.
External Supermemory Scoped profile/search/remember/forget operations plus optional mirrored memory history.

CoWork treats Supermemory results as soft context:

useful prior context
lower priority than the current user message
separate from the local workspace kit and local archive memory

Prompt Injection

If Inject Supermemory Profile Into Prompts is enabled, CoWork performs a scoped profile fetch during prompt construction and injects:

static facts
dynamic recent context
a small set of relevant external memories for the current task/query

This happens in:

chat turns
execution turns
follow-up turns

The injected block is wrapped as pinned profile-style context, but it remains advisory. If the user gives newer or conflicting information in the current conversation, the current message should win.

Mirroring Behavior

If Mirror Memory Writes is enabled, CoWork mirrors non-private archive-memory captures into Supermemory.

Current mirroring source:

MemoryService.capture(...) for non-private memory entries

Current mirrored payload shape:

raw memory content as an external document
metadata including workspace ID, workspace name, task ID, memory type, and capture timestamp

Current exclusions:

private/strict-mode memory entries are not mirrored
clipboard-only/private sensitive content remains local
this integration does not currently stream every chat turn into Supermemory conversations

That last point matters: CoWork currently mirrors memory captures, not the full conversation transcript lifecycle.

Tooling

`supermemory_profile`

Fetches the current profile for the workspace-scoped container, optionally with a query.

Use it when the agent needs:

durable user context
recent external context
relevant external memories before answering

`supermemory_search`

Runs explicit search against the configured container.

Configurable options include:

query
containerTag
limit
threshold
rerank
searchMode

Supported search modes in CoWork:

hybrid
memories

`supermemory_remember`

Creates a durable external memory directly in Supermemory.

Use it for:

preferences
project facts
stable context worth keeping outside the local machine

`supermemory_forget`

Forgets an external memory by:

exact memory ID, or
exact content text

Use it when external memory is outdated or incorrect.

Container Tags And Scope

By default, CoWork resolves one workspace-scoped container tag from the configured template.

Examples:

cowork:workspace-123
cowork:Client-A

Custom container entries can also be stored in the UI as named namespaces. Today, those entries are:

configuration metadata
useful for human/operator reference
usable with explicit containerTag overrides on the Supermemory tools

What CoWork does not do yet:

automatic model-driven container switching
heuristic auto-routing between work/personal/project containers

If you want multi-container routing today, use explicit containerTag values with the Supermemory tools.

Failure Handling

Supermemory failures should not brick the CoWork runtime.

Current safeguards:

short request timeout
best-effort prompt injection
best-effort background mirroring
circuit breaker after repeated request failures

When the circuit breaker opens, CoWork pauses Supermemory requests temporarily and keeps running with local memory only.

The Memory Hub shows:

whether the API key is configured
the latest connection-test result
the most recent provider error
circuit-breaker pause state when active

Privacy And Boundaries

Important boundaries:

CoWork's local memory remains the primary durable memory system
Supermemory is optional and external
mirrored local memory writes can leave the device
private memory entries are not mirrored
workspace kit files remain local and governed by CoWork's existing memory/runtime policies

If you want fully local-only operation, leave Supermemory disabled.

Subconscious Loop

Supervisor Mode Discord

Was this page helpful?Edit this page on GitHub