Agent Skills: Memory Exploration (RLM-style)

Memory Exploration (RLM-style)

Instead of injecting top-k memories, explore the memory graph dynamically.

How It Works

1. Start with a query
2. Get initial hints via explore_recall
3. Iteratively decide: peek, expand, follow neighbors, or answer
4. Accumulate relevant findings
5. Answer when sufficient context gathered

Exploration Protocol

You have these primitives (via chitta RPC):

| Tool | Purpose | Token Cost | |------|---------|------------| | explore_recall | Semantic search, returns hints (id, title, score) | ~100 | | explore_peek | Get 200-char summary of a memory | ~50 | | explore_expand | Get full memory content | ~200-500 | | explore_neighbors | Get triplet connections from a node | ~100 |

Agent Loop

query = user's question
context = []
trace = []
max_iterations = 10

# Initial hints
hints = explore_recall(query, limit=5)
trace.append(("recall", query, hints))

for i in range(max_iterations):
    # Decide next action based on query + current context
    action = decide_action(query, context, hints)

    if action == "ANSWER":
        break
    elif action.startswith("PEEK"):
        id = extract_id(action)
        summary = explore_peek(id)
        context.append(summary)
        trace.append(("peek", id, summary))
    elif action.startswith("EXPAND"):
        id = extract_id(action)
        full = explore_expand(id)
        context.append(full)
        trace.append(("expand", id, len(full)))
    elif action.startswith("NEIGHBORS"):
        node = extract_node(action)
        neighbors = explore_neighbors(node)
        hints.extend(relevant_neighbors(neighbors))
        trace.append(("neighbors", node, len(neighbors)))
    elif action.startswith("RECALL"):
        new_query = extract_query(action)
        new_hints = explore_recall(new_query, limit=5)
        hints.extend(new_hints)
        trace.append(("recall", new_query, new_hints))

# Generate answer from accumulated context
answer = synthesize(query, context)
return answer, trace

Decision Prompt

At each iteration, decide:

Query: {query}

Current context ({len(context)} items):
{format_context(context)}

Available hints:
{format_hints(hints)}

Recent trace:
{format_trace(trace[-3:])}

What should I explore next?
- PEEK(id) - get summary of a promising hint
- EXPAND(id) - get full content (use sparingly)
- NEIGHBORS(node) - explore triplet connections
- RECALL("query") - search for more hints
- ANSWER - I have enough context

Respond with just the action, e.g.: PEEK(00000000-0000-0000-0000-0000000001bd)

When to Use

Use /explore instead of regular recall when:

• Memory store is large (>1000 memories)
• Query needs associative exploration
• Top-k might miss relevant context
• You want to understand the exploration path

Execution

When user invokes /explore <query>:

1. Initial recall: Call explore_recall with the query
2. Show hints: Display the hints found
3. Exploration loop (max 10 iterations):
   • Based on hints and accumulated context, decide:
     • PEEK(id) if a hint looks promising but need more info
     • EXPAND(id) if confident this memory is highly relevant
     • NEIGHBORS(node) if you want to follow relationships
     • RECALL("new query") if you need different search terms
     • ANSWER if you have enough context
   • Execute the action using the chitta RPC tool
   • Add results to context
4. Synthesize: Generate answer from accumulated context
5. Show trace: Display exploration path (what was searched, peeked, expanded)

Token Comparison

Show at the end:

• Tokens used in exploration: ~N (sum of all actions)
• Equivalent full_resonate: ~M (10 full memories)
• Savings: X%

Example

User: /explore How does the daemon handle authentication?

Claude: Starting exploration for "How does the daemon handle authentication?"

[Step 1] explore_recall("daemon authentication")
  Found 5 hints:
  [45%] id:123 - [mcp] cc-soul socket connection
  [38%] id:456 - [wisdom] JSON-RPC protocol
  [32%] id:789 - Unix socket permissions
  ...

[Step 2] PEEK(123) - socket connection looks relevant
  Summary: "cc-soul daemon uses Unix socket at /tmp/chitta-*.sock..."

[Step 3] NEIGHBORS("daemon")
  Found: daemon → uses → json_rpc
         daemon → listens_on → unix_socket

[Step 4] EXPAND(456) - JSON-RPC protocol details needed
  Full content loaded (342 chars)

[Step 5] ANSWER - have enough context

=== Answer ===
The daemon uses Unix socket authentication via file permissions.
No explicit auth - relies on filesystem access control to /tmp/chitta-*.sock.

=== Exploration Trace ===
1. recall("daemon authentication") → 5 hints
2. peek(123) → 200 chars
3. neighbors("daemon") → 5 connections
4. expand(456) → 342 chars

Tokens: ~692 exploration vs ~1500 full_resonate (54% savings)

RPC Commands

Use these chitta commands during exploration:

# Lightweight recall
chitta explore_recall --query "your search" --limit 5

# Peek at summary
chitta explore_peek --id "00000000-0000-0000-0000-000000000123"

# Full content
chitta explore_expand --id "00000000-0000-0000-0000-000000000123"

# Graph neighbors
chitta explore_neighbors --node "Mind" --direction both