Mempalace as session db #27710

Activity

1. github-actions

   assigned

   kitlangton

   on May 15, 2026
2. github-actions commented on May 15, 2026

   

   github-actionsBot

   on May 15, 2026 – with GitHub Actions

   Contributor

   More actions

   This issue might be a duplicate of existing issues. Please check:

   • [FEATURE]:Context-Aware Proxy Middleware: intercept requests before LLM to reduce token costs 96–99% via MemPalace + SocratiCode #26685: Similar request involving MemPalace for context/token reduction across sessions
3. geco commented on Jun 3, 2026

   

   geco

   on Jun 3, 2026

   More actions

   Hey @d0uub, I built an OpenCode plugin that persists conversations to MemPalace in real-time — might be what you are looking for:

   https://github.com/geco/opencode-mempalace-persistence

   Happy to collaborate or discuss how this could be extended as a native session backend. Also just opened a PR to list it on the ecosystem page: #30540
4. d0uub commented on Jun 4, 2026

   

   d0uub

   on Jun 4, 2026

   Author

   More actions

   Thanks @geco for sharing your persistence plugin! It's great to see the MemPalace integration evolving.

   I wanted to share another approach that complements the persistence work - automatic Memory Injection with Wake-up Context using the experimental.chat.messages.transform hook.

   While your plugin handles the write side (saving conversations), I've been experimenting with the read side - automatically injecting relevant memories and wake-up context before the AI processes each message, without relying on model discipline to search MemPalace.

   Key Implementation

   The core mechanism uses experimental.chat.messages.transform to intercept and augment user messages:

   "experimental.chat.messages.transform": async (_input, output) => {
     if (!output?.messages?.length) return output
     const lastUser = [...output.messages].reverse().find(m => m.info?.role === "user")
     if (!lastUser) return output
   
     // 1. Wake-up: Inject L0+L1 context on first message
     if (!wakeupDone) {
       wakeupDone = true
       const ctx = wakeup()  // Calls `mempalace wake-up`
       if (ctx) {
         lastUser.parts.push({
           id: `mp-wakeup-${Date.now()}`,
           type: "text",
           synthetic: true,
           text: `[MemPalace Context]\n${ctx}\n[/MemPalace Context]`
         })
       }
     }
   
     // 2. Recall: Search and inject relevant memories
     const query = textFromParts(lastUser.parts, true)
     const memories = search(query)  // Similarity search with budget control
     if (memories) {
       lastUser.parts.push({
         id: `mp-recall-${Date.now()}`,
         type: "text",
         synthetic: true,
         text: `[MemPalace Recall]\n${memories}\n[/MemPalace Recall]`
       })
     }
     
     return output
   }

   Benefits

   1. Guaranteed Recall - Doesn't depend on model following AGENTS.md instructions (which smaller models skip ~60% of the time)
   2. Wake-up Context - Provides L0+L1 identity/project context automatically, similar to AAAK pattern
   3. Token Efficient - Uses similarity thresholds (MIN_SIMILARITY=0.20) and character budgets (MAX_INJECT_CHARS=900)
   4. Zero Configuration - No need for separate AGENTS.md or identity.txt files

   Storage Side

   For the persistence side, maybe consider following:

   function store(content) {
     // Write to temp file
     const tempFile = `/tmp/mp-store-${Date.now()}.txt`
     writeFileSync(tempFile, content)
     
     // Call mempalace mine
     execSync(`mempalace mine ${tempFile} --mode convos`)
     
     // Cleanup
     unlinkSync(tempFile)
   }
5. d0uub

   closed this as completedon Jun 4, 2026
6. geco commented on Jun 4, 2026

   

   geco

   on Jun 4, 2026

   More actions

   @d0uub I ended up implementing this as a plugin — opencode-mempalace-persistence. It uses experimental.chat.messages.transform to inject relevant MemPalace memories into every prompt automatically (the approach you pioneered). The plugin also saves conversations back to MemPalace in real-time.

   Repo: https://github.com/geco/opencode-mempalace-persistence
   npm: opencode-mempalace-persistence@1.3.1

   Thanks for the inspiration — your comment sparked the auto-inject approach.