memory-exploring

Exploring Codebases with hoangsa-memory

Understand unfamiliar code without reading every file. The hybrid retriever (memory_recall) finds entry points by intent; the graph (memory_symbol_context) walks relationships. For whole-shape questions the graph answers directly — reach for these INSTEAD of grepping:

• "What are the main components?" → memory_graph_communities() — clusters of tightly-coupled symbols (the architecture map).
• "Walk me through the main flow" / "trace execution" → memory_graph_processes() — flows from entry points (::main) down the call graph.
• "How does A reach B?" → memory_graph_paths({from, to}) — shortest dependency path.
• "Everything connected to X" → memory_graph_query({start}) — traverse callers/callees/refs/imports to any depth.
• "Can untrusted input reach a dangerous call?" → memory_taint_paths({sources, sinks}) — source→sink dataflow (index with --pdg).
• "Who publishes / subscribes to this event?" → memory_event_trace({topic}) — pub/sub flow the call graph can't connect (decoupled by a broker).

A single one of these usually replaces a dozen Greps.

When to Use

• "How does authentication work?"
• "What's the overall architecture?"
• "Show me the main components"
• "Where is the database layer?"
• Onboarding to an unfamiliar repo.

Workflow

1. resources/read hoangsa-memory://memory/MEMORY.md           → durable facts first
2. memory_recall({query: "<concept>"})                 → candidate chunks
3. memory_symbol_context({fqn})                        → 360° around each
   (architecture-wide? → memory_graph_communities / _processes instead of step 2)
4. Walk callers to find entry points                  → root of flow
   (or memory_graph_query/_paths — one call vs many manual hops)
5. Cite chunk ids in the explanation

1. Read durable memory

Before recalling, skim MEMORY.md. Facts there encode architectural decisions and conventions you'd otherwise miss. If a fact names the module you're about to explore, it saves a recall.

resources/read { uri: "hoangsa-memory://memory/MEMORY.md" }
resources/read { uri: "hoangsa-memory://memory/LESSONS.md" }

(Or equivalently: memory_show.)

2. Recall by intent

Use the user's own vocabulary. For "how does auth work", try:

memory_recall { query: "authentication login session token" }

Prefer nouns over verbs. The hybrid retriever rewards topical density. If the first query misses, widen with synonyms or narrow with concrete symbols you see in the README.

3. Drill with symbol_context

Pick the top-ranked chunk's FQN. Then:

memory_symbol_context { fqn: "auth::verify_token", limit: 16 }

Sections to read:

• callers — who invokes this → likely entry points.
• callees — what it delegates to → next layer.
• siblings — other symbols in the same file → adjacent concerns.
• references — non-call uses (type refs, trait bounds).
• extends / extended_by — inheritance / impl relationships.

4. Walk upward to the entry point

The entry point of a flow usually has few callers (it's a request handler, a CLI command, a main loop). Repeatedly follow callers until the list collapses to 0–1. That symbol is the flow's root.

Once you have root + N inner symbols, the story writes itself:

Auth starts at router::handle_login (1 caller: main::serve). It calls auth::verify_credentials → db::users::find_by_email → password::verify_hash. On success, auth::issue_token signs a JWT via crypto::sign. All five symbols live under crates/auth/.

5. Cite chunks

Every claim about behaviour should cite the chunk id from step 2 or a FQN+line from step 3. Ungrounded claims about "what the code does" are the #1 source of drift in exploratory sessions.

Anti-patterns

• Skimming the file tree instead of recalling. Names lie. The retriever ranks by semantic + lexical + graph signals; directory names are only one of those.
• Ignoring MEMORY.md. If a fact says "auth lives in crates/net, not crates/auth — historical accident", you save 20 minutes of misdirected recall.
• Stopping at the first chunk. The top-ranked hit is the most similar chunk, not necessarily the most important. Read the top 3–5 before deciding where to drill.
• Recalling on full sentences. "How does the system authenticate users?" is noisier than authentication login session. Strip function words.

When recall returns nothing

If memory_recall gives (no matches — did you run memory_index?), the graph is empty. Run hoangsa-memory index . (CLI) or memory_index (MCP) and retry. If it still returns nothing, the concept genuinely isn't indexed — ask the user for a file path to anchor from, or fall back to Grep for a literal keyword.