🎉 History Extension: Summary Generation and 🧠 Long-term Memory with VectorCode

[ravitemer]

Hi guys!

I'm excited to share an update to the codecompanion-history.nvim extension - we now have summary generation and long-term memory! Thanks to @Davidyz for pitching this at ravitemer/codecompanion-history.nvim#16

🚀 How It Works

Basic Chat Management

Just like before, your CodeCompanion chats are automatically saved (or manually with sc). Browse your saved chats with gh and select any chat to restore it completely with all context, tools, and references intact.

Summary Generation Workflow

When your conversation gets too long and approaches context limits:

1. Press gcs (create summary) - make sure to configure your preferred adapter/model to avoid premium models
2. Visual feedback - chat title shows "🔄 Generating summary..." then "📝" when complete (also shown in while browsing saved chats that have been summarized)
3. Summary saved - intelligent summary is generated and stored as markdown You can customize the prompt for generating these summaries if you want. See below.

Using Summaries

When you want to reference previous conversations:

1. Press gbs (browse summaries) to see all your saved summaries
2. Select a summary to add it as a reference to your current chat
3. Context added - appears as > - <summary>chat_title</summary> in your chat

Memory Tool Integration

Special thanks to @Davidyz for the fantastic VectorCode CLI integration and implementing the @memory tool that brings true semantic search to our conversations

If you have VectorCode cli available in the PATH:

• Summaries are automatically indexed for semantic search
• Use the @memory tool in new chats - your LLM can now search through previous conversations
• Ask things like "What did we discuss about error handling?" and the LLM will find relevant past conversations

⚙️ Configuration

require("codecompanion").setup({
    extensions = {
        history = {
            enabled = true,
            opts = {
                -- Summary system
                summary = {
                    create_summary_keymap = "gcs", -- Generate summary for current chat
                    browse_summaries_keymap = "gbs", -- Browse saved summaries
                    generation_opts = {
                        adapter = nil, -- e.g "copilot"  (nil = current chat adapter)
                        model = nil, -- e.g "gpt-4.1" (nil = current chat model)
                        context_size = 90000, -- Max tokens that the model supports 
                        include_references = true, -- Include slash command content
                        include_tool_outputs = true, -- Include tool execution results
                        system_prompt = nil, -- Custom system prompt (string or function)
                        format_summary = nil -- Custom function to format generated summary (e.g removing <think/> tags)
                    },
                },
                
                -- Memory system (requires VectorCode CLI)
                memory = {
                    auto_create_memories_on_summary_generation = true, -- Index summaries automatically
                    vectorcode_exe = "vectorcode", -- Path to VectorCode executable
                    tool_opts = { default_num = 10 }, -- Default number of memories to retrieve
                    notify = true, -- Show notifications for indexing progress
                    index_on_startup = false, -- Index all existing summaries on startup
                },
               }
        }
    }
})

Check out codecompanion-history.nvim for full documentation

Warning

Title and summary generation defaults to current chat's adapter and model. Make sure to set cheaper models in title_generation_opts and summary.generation_opts to avoid using premium models.

🔮 Future Roadmap

While this implementation provides manual summary generation and sort of long term memory, future enhancements could include:

• Automatic Summary Generation: Generate summaries when conversations reach certain length thresholds
• Dynamic Context Management: Replace old messages with summaries in long conversations to stay within context limits
• Summary Refinement: Edit and improve summaries over time

[jinzhongjia]

Great, this is an alternative iteration or inheritance implementation. I tried to implement a pr before, but it was not accepted, which is very useful, thanks for your work

[jinzhongjia]

We can adopt a method like gemini cli or vscode copy chat to automatically generate summary, and their open source code is worth learning