Skip to content

Automatic Context Injection: Giving Agents "Foresight"

What You'll Learn

By the end of this lesson, you will be able to:

  1. Understand why the Agent knows your coding habits and project architecture from the very beginning.
  2. Master the "3D Model" of context injection (User Profile, Project Knowledge, Relevant Memories).
  3. Learn to use keywords (like "Remember this") to actively intervene in the Agent's memory behavior.
  4. Configure the number of injected items to balance context length with information richness.

Core Concept

Before memory plugins, every time you started a new session, the Agent was a blank slate. You had to repeatedly tell it: "I use TypeScript," "This project uses Next.js."

Context Injection solves this problem. It's like slipping a "mission briefing" into the Agent's mind the moment it wakes up.

Trigger Timing

opencode-supermemory is extremely restrained, triggering automatic injection only on the first message of the session.

  • Why only the first? Because this is the critical moment to establish the session's tone.
  • What about subsequent messages? Subsequent messages are not automatically injected to avoid disrupting the flow of conversation, unless you actively trigger it (see "Keyword Trigger" below).

The 3D Injection Model

The plugin fetches three types of data in parallel and combines them into a [SUPERMEMORY] prompt block:

Data DimensionSourceFunctionExample
1. User Profile (Profile)getProfileYour long-term preferences"User likes functional programming", "Prefers arrow functions"
2. Project Knowledge (Project)listMemoriesGlobal knowledge of the current project"This project uses Clean Architecture", "APIs are in src/api"
3. Relevant Memories (Relevant)searchMemoriesPast experiences related to your first sentenceYou ask "How to fix this bug", it finds similar fix records from the past

What is Injected?

When you send your first message in OpenCode, the plugin silently inserts the following content into the System Prompt in the background.

Click to view the actual structure of injected content
text
[SUPERMEMORY]

User Profile:
- User prefers concise responses
- User uses Zod for all validations

Recent Context:
- Working on auth module refactoring

Project Knowledge:
- [100%] Architecture follows MVC pattern
- [100%] Use 'npm run test' for testing

Relevant Memories:
- [85%] Previous fix for hydration error: use useEffect

After seeing this information, the Agent behaves like a veteran employee who has worked on this project for a long time, rather than a new intern.


Keyword Trigger Mechanism (Nudge)

In addition to the automatic injection at the start, you can "wake up" the memory function at any time during the conversation.

The plugin has a built-in Keyword Detector. As long as your message contains specific trigger words, the plugin sends an "invisible hint" (Nudge) to the Agent, forcing it to call the save tool.

Default Keywords

  • remember
  • save this
  • don't forget
  • memorize
  • take note
  • ... (See source code configuration for more)

Interaction Example

You type:

The API response format has changed here, remember to use data.result instead of data.payload from now on.

Plugin detects "remember":

(Background injection hint): [MEMORY TRIGGER DETECTED] The user wants you to remember something...

Agent reacts:

Understood. I will remember this change. (Calls supermemory.add in the background to save memory)


Advanced Configuration

You can adjust the injection behavior by modifying ~/.config/opencode/supermemory.jsonc.

Common Configuration Options

jsonc
{
  // Whether to inject user profile (default true)
  "injectProfile": true,

  // How many project memories to inject each time (default 10)
  // Increasing this helps the Agent understand the project better but consumes more Tokens
  "maxProjectMemories": 10,

  // How many user profile items to inject each time (default 5)
  "maxProfileItems": 5,

  // Custom trigger words (supports Regex)
  "keywordPatterns": [
    "note this",
    "save forever"
  ]
}

Tip

After modifying the configuration, you need to restart OpenCode or reload the plugin for changes to take effect.


FAQ

Q: Does the injected information consume a lot of Tokens?

A: It consumes some, but it is usually manageable. Under the default configuration (10 project memories + 5 profile items), it takes up about 500-1000 Tokens. For the 200k context of modern large models (like Claude 3.5 Sonnet), this is a drop in the bucket.

Q: Why didn't it react when I said "remember"?

A:

  1. Check if the spelling is correct (Regex matching is supported).
  2. Confirm if the API Key is configured correctly (it won't trigger if the plugin is not initialized).
  3. The Agent might have decided to ignore it (although the plugin forces a hint, the Agent has the final say).

Q: How are "Relevant Memories" found?

A: It performs a semantic search based on the content of your first message. If your first sentence is just "Hi", it might not find any useful relevant memories, but "Project Knowledge" and "User Profile" will still be injected.


Summary

  • Automatic Injection triggers only on the first message of the session.
  • The 3D Model includes User Profile, Project Knowledge, and Relevant Memories.
  • Keyword Trigger allows you to command the Agent to save memories at any time.
  • You can control the amount of injected information via the Configuration File.

Next Step

Next, we will learn Tool Set Details: Teaching Agents to Remember.

You will learn:

  • How to manually use tools like add, search, etc.
  • How to view and delete incorrect memories.

Appendix: Source Code Reference

Click to expand source code locations

Updated: 2026-01-23

FeatureFile PathLine Number
Injection Trigger Logicsrc/index.ts125-176
Keyword Detectionsrc/index.ts34-37
Prompt Formattingsrc/services/context.ts14-64
Default Configurationsrc/config.ts44-54

Key Functions:

  • formatContextForPrompt(): Assembles the [SUPERMEMORY] text block.
  • detectMemoryKeyword(): Regex matches trigger words in user messages.

Next Step

Next, we will learn Tool Set Details: Teaching Agents to Remember.

You will learn:

  • Master 5 core tool modes including add, search, profile
  • How to manually intervene and correct Agent memories
  • Triggering memory saves using natural language