remember

Recuerd0 Memory

Preserve, version, and organize knowledge from AI conversations using the Recuerd0 CLI (recuerd0). Run commands via Bash and interpret the structured JSON output to manage the user's workspaces and memories. The full command catalog, flags, and output format live in references/cli-reference.md — read it for exact syntax.

Operate proactively: watch the conversation for capture-worthy moments and act on them without waiting to be asked. Operate with discipline: write no duplicate memory, always route to the right workspace, and announce every save in one line so the user knows it happened.

All memory content MUST be Markdown. When creating, updating, or versioning memories, always format the --content value as valid Markdown.

---

When to Capture

Before capturing anything, apply this gate. It is the single rule that decides whether a candidate earns a memory:

Only persist a discovery if it would not be obvious from reading the code, contradicts a sane default, or cost real time to find; before creating, search for an existing memory on the topic and version it instead.

Operationally:

• Not obvious from the code — if a competent developer reading the source would arrive at the same conclusion in under a minute, do not persist it. Memories are for what the code doesn't tell you: the reason behind a choice, the gotcha that wasted an hour, the constraint that isn't written down.
• Contradicts a sane default — capture the surprises. "We deviate from the framework default here, because X" is worth a memory. "We follow the framework default" is not.
• Cost real time to find — if it was a fast lookup, it'll be a fast lookup next time. If it took debugging, bisecting, or reading three files to understand, persist it so it never costs that again.
• Version, don't duplicate — running the search step (Dedup-Before-Write Protocol) before every write is mandatory, not optional. The default action on a topic match is memory version create, not memory create.

If a candidate fails all three tests, skip it silently — do not announce a non-save.

Capture proactively when any of these signals appear in the conversation:

• A decision was made — architecture choice, library choice, tradeoff with stated reasoning ("we picked Postgres because…")
• A non-obvious bug was diagnosed — root cause is now understood, not just symptoms
• The user stated a strong preference — "I always want X", "never do Y", "from now on…"
• A breakthrough discovery — a pattern, library quirk, workaround, or gotcha worth remembering
• End of a focused session — natural stopping point on a specific topic, before context shifts

Do NOT capture when:

• The conversation is exploratory or chitchat
• Work is in progress — wait until the decision/fix actually lands
• The information is already in the project's CLAUDE.md or AGENTS.md
• The workspace already contains it (the dedup protocol below catches this)
• The user explicitly says not to remember something

When in doubt, capture. A duplicate prevented by dedup is cheap; a lost decision is expensive.

---

Workspace Routing

Always resolve the target workspace before asking the user. Follow this decision tree:

1. Identify the project — run pwd and git config --get remote.origin.url. The current working directory and git remote are the strongest signals.
2. Check project-local config — the CLI walks parent directories looking for .recuerd0.yaml. If you can run any recuerd0 command without --workspace and it succeeds, the resolution worked.
3. Match by name — run recuerd0 workspace list --pretty and look for a workspace whose name matches the project directory or repo name.
4. Single clear match — use it without asking. Mention it in your one-line save notice so the user can correct if wrong.
5. Multiple plausible matches — pick the closest and confirm in one line ("Saving to workspace 'rails-patterns' (3 candidates) — say so if wrong").
6. No match at all — create a new workspace with a name derived from the project: recuerd0 workspace create --name "<project-name>" --description "<inferred from README or git remote>". You may create without asking. Mention it in the save notice.
7. Never silently dump into a generic default workspace. Every save must land somewhere semantically correct.

---

Dedup-Before-Write Protocol

Before any memory create, you MUST:

1. Load the workspace context — recuerd0 workspace context <id> --pretty. This returns the workspace metadata plus the user's pinned memories filtered to this workspace, in one call. Pinned memories are the most likely candidates for an update vs. a duplicate.

2. Search for related memories (multi-query). A single search query finds a memory only when its wording overlaps the recorded wording. Concepts recorded as "boot" won't surface for a query about "startup"; "broadcast" won't surface for "realtime updates". To avoid creating a duplicate of a memory you simply failed to find, search with several phrasings, not one.

   Generate 3–5 query variants for the new content, then run a search for each and union the results before deciding create vs. version vs. update. Construct the variants to span vocabulary, not just restate the title:

   • One variant from the recorded jargon — the exact technical terms the memory likely uses (broadcast_refresh_to, Zeitwerk, script_name).
   • One variant from intent — how a developer who doesn't know the jargon would ask ("send updates to one user only", "autoload fails at boot", "url prefix per account").
   • One variant swapping synonyms on the key noun/verb — boot↔startup↔initialization, broadcast↔stream↔push↔realtime, middleware↔rack, locale↔language↔i18n.
   • Optionally one tag-scoped or category-scoped pass if the topic maps to a known tag.

   Run each variant intra-workspace, and run at least one variant across all workspaces (no --workspace) for the cross-workspace link check:

   recuerd0 search "broadcast_refresh_to per user" --workspace <id> --pretty
   recuerd0 search "send turbo updates to one user" --workspace <id> --pretty
   recuerd0 search "scope realtime stream single account" --workspace <id> --pretty
   recuerd0 search "broadcast_refresh_to per user" --pretty   # cross-workspace
   

   Union the hits by memory id (a memory found by any variant counts as found). Then proceed to the decision in step 4 against the unioned candidate set.

   FTS5 operator hygiene — do not let query words become operators:

   • and, or, not are FTS5 operators. Never include them as plain search words. "graphics not rendering" is parsed as graphics NOT rendering and will silently exclude matches. Drop them or quote the phrase.
   • Quote multi-word phrases that must match as a unit: '"access request"'.
   • Keep each variant to 2–4 distinctive terms. Long natural-language sentences fail because every token must match.

   Stop condition. If two or more variants return zero results and one returns a weak/unrelated hit, treat the topic as not found and create — but log in your reasoning that recall was thin, so a later audit can catch a possible twin. If any variant returns a strong same-topic match, default to memory version create.

3. Search across workspaces — also run recuerd0 search "<key terms>" --pretty without --workspace to find related memories in other workspaces in the same account. If a strong cross-workspace match exists, after you've created or versioned the new memory, ask the user in one line: Link this to memory <id> '<title>' in workspace <name>? (y/n). On yes, run recuerd0 memory link add <new_id> --to <other_id>. On anything else, skip the link. Never link silently.

4. Decide based on what you find:

   • Strong match (same topic, same scope, just evolved): use recuerd0 memory version create --workspace <id> <memory_id> --content - to add a new version. Default to versioning — recuerd0's whole versioning model exists for this. Preserve history.
   • Wrong match (the existing memory is incorrect, not just outdated): use recuerd0 memory update --workspace <id> <memory_id> to overwrite.
   • No match: only then call recuerd0 memory create.

   If a candidate "duplicate" memory is large, use recuerd0 memory read grep <id> "<key term>" to confirm the fact already exists before loading the full body — this is much cheaper than memory show for long memories.

5. Pick a category from the four values in the Categories section above (decision, discovery, preference, general). Pass it via --category on create or version create.

Examples

Versioning an evolving decision:

recuerd0 workspace context 1 --pretty
recuerd0 search "auth strategy" --workspace 1 --pretty
# Finds memory 42: "Auth strategy" — outdated
recuerd0 memory version create --workspace 1 42 \
  --category decision \
  --content - <<'EOF'
# Auth strategy v2

We migrated from session-only to session + bearer-token API auth.
Reason: needed programmatic API access for the CLI.
EOF

Fresh capture after dedup miss:

recuerd0 workspace context 1 --pretty
recuerd0 search "fts5 tokenizer" --workspace 1 --pretty
# No matches
recuerd0 memory create --workspace 1 \
  --title "FTS5 trigram tokenizer for substring search" \
  --tags "sqlite,fts5,search" \
  --source "claude-code-session" \
  --category discovery \
  --content - <<'EOF'
# FTS5 trigram tokenizer

Substring matching in FTS5 requires the trigram tokenizer…
EOF

Cross-workspace dedup hit with confirmed link:

Working in the mobile-app workspace (id 3), capturing a decision about how the mobile client refreshes OAuth tokens. The intra-workspace search finds nothing, but the cross-workspace search surfaces a clearly related memory in rails-app.

recuerd0 workspace context 3 --pretty
recuerd0 search "oauth refresh token" --workspace 3 --pretty
# No matches in workspace 3
recuerd0 search "oauth refresh token" --pretty
# Strong match: memory 42 "Auth strategy" in workspace rails-app (id 1)
recuerd0 memory create --workspace 3 \
  --title "OAuth refresh on the mobile client" \
  --tags "oauth,mobile,auth" \
  --source "claude-code-session" \
  --category decision \
  --content - <<'EOF'
# OAuth refresh on the mobile client

The mobile client refreshes its OAuth access token in the background…
EOF
# Server returns the new memory id: 118

Then ask the user in one line:

Link this to memory 42 "Auth strategy" in workspace rails-app? (y/n)

On y:

recuerd0 memory link add 118 --to 42

Final save notice:

✓ Saved to workspace mobile-app (id 3) as "OAuth refresh on the mobile client" [decision] (id 118) [created]
  → linked to "Auth strategy" in workspace rails-app (id 42)

---

Categories

Every memory carries a category, picked from a locked four-value enum. The server defaults new memories to general when no category is sent, but you should always pass --category explicitly on every memory create and memory version create so the choice is deliberate.

Value	Label	When to pick it
decision	Decision	Architecture choices, library picks, tradeoffs with stated reasoning
discovery	Discovery	Non-obvious findings — gotchas, root causes, patterns, library quirks
preference	Preference	User-stated rules ("always X", "never Y")
general	General	Catch-all and default

Picking heuristic: if torn between two, prefer the more specific one — discovery over general, decision over discovery. general is a fallback, not a default.

Versions inherit the parent memory's category unless you explicitly override with --category on memory version create.

---

Memory Links

Memory links — sometimes called tunnels — are undirected, unlabeled "see also" connections between two memories within the same account. Unlike tags or workspaces, links cross workspace boundaries: an "auth strategy" memory in the rails-app workspace can be linked to "auth strategy" in the mobile-app workspace, letting you express that memories in two different projects cover related territory.

Key constraints:

• Same account only — both memories must belong to the current account. Cross-account links are rejected by the server.
• Undirected — linking A↔B once is enough. You do not need (and cannot create) a reverse link. The server dedupes A→B and B→A.
• No labels, no metadata — a link is just a connection. There is no relationship type, no description, no weight.
• No self-links — a memory cannot link to itself.
• Both endpoints must exist at link time.
• No automatic linking — create a link only after explicit user confirmation.

Commands:

recuerd0 memory link list <memory_id> [--workspace ID]                    # List all links for a memory
recuerd0 memory link add <memory_id> --to <other_memory_id> [--workspace ID]    # Create a link
recuerd0 memory link remove <memory_id> --to <other_memory_id> [--workspace ID] # Remove a link

The --workspace flag refers to the source memory's workspace; the target memory may live in any workspace within the same account. The memory show JSON and the workspace context JSON now include a links_count field on each memory so you can see at a glance how connected something is.

Confirm before linking — always. Your job is to suggest a link when a strong cross-workspace match surfaces during dedup. Ask the user in one line and wait for a one-word approval. Create a link only with that approval. A noisy link graph is worse than no graph.

When to link

Link when:

• The new memory closely echoes a memory in another workspace — same topic, different project (e.g. "JWT refresh strategy" exists in both the API and mobile workspaces).
• The user explicitly says "this is like that other thing in workspace X" or otherwise points at a cross-workspace relationship.
• The cross-workspace dedup search surfaces a strong match — same or near-identical title, clearly the same subject, just scoped to a different project.

DO NOT link when:

• The connection is weak or speculative ("both about Rails" is too broad — that describes hundreds of memories).
• The other memory lives in an archived workspace.
• The two memories are already linked (check recuerd0 memory link list first if unsure).
• The user has not confirmed. No confirmation, no link.

The default is not to link. Linking is the exception, reserved for genuine cross-project topic matches that a user would actually want to traverse.

---

Save Notice Convention

Every successful capture must produce one line of user-facing output, in this format:

✓ Saved to workspace <name> (id <ws_id>) as "<title>" [<category>] (id <mem_id>) [<action>]
  → linked to "<other_title>" in workspace <other_name> (id <other_id>)

Where <action> is one of: created, versioned, updated. The second indented line is emitted only when a link was actually created in this capture. If no link was created, omit the link line entirely — do not say "no links". Examples:

✓ Saved to workspace rails-patterns (id 1) as "FTS5 trigram tokenizer for substring search" [discovery] (id 87) [created]

✓ Saved to workspace mobile-app (id 3) as "OAuth refresh on the mobile client" [decision] (id 118) [created]
  → linked to "Auth strategy" in workspace rails-app (id 42)

✓ Saved to workspace rails-patterns (id 1) as "Auth strategy" [decision] (id 42) [versioned]

Do not narrate the dedup process, the search results, or the workspace resolution unless something went wrong or the user asked. The one-liner is enough.

---

CLI Reference

The command catalog, output format, global flags, breadcrumbs, exit codes, and usage patterns live in references/cli-reference.md. Every command emits structured JSON — check success, read data, follow the suggested breadcrumbs, and pass --pretty when showing output to the user.

---

Save Session as Memory

When the user explicitly asks to "save this session", or when you reach the end of a focused working session and detect capture-worthy content (see When to Capture), produce a curated memory.

Steps

1. Resolve the workspace following the Workspace Routing decision tree above.
2. Run the dedup protocol above. Decide whether you're creating, versioning, or updating.
3. Generate a transcript in Markdown from the current conversation context. All memory content MUST be Markdown. Use this structure:

# <Conclusion-first title>

## Goal
What the user set out to accomplish.

## Summary
2-3 paragraph overview of what happened, decisions made, and outcomes.

## Key Changes
- Bullet list of files created, modified, or deleted with brief descriptions

## Decisions & Rationale
- Important choices made and why

## Learnings
- Patterns, gotchas, or insights worth preserving

4. Format guardrails:

   • Title: declarative, scoped, ≤80 chars. Lead with the conclusion ("FTS5 errors surface as ActiveRecord::StatementInvalid"), not the topic ("FTS5 stuff").
   • Tags: 3–6, lowercase, hyphenated. Aim for one domain tag (auth, deploy), one tech tag (rails, sqlite), one type tag (decision, pattern, bugfix).
   • Source: claude-code-session for proactive captures, manual for explicit user requests, <project>-decision for architecture-decision memories.
   • Category: required. Pick from decision, discovery, preference, general. Lean toward the more specific choice — general is a fallback, not a default.

5. Save via the CLI by piping the transcript through stdin:

cat <<'TRANSCRIPT' | recuerd0 memory create --workspace <id> --title "<title>" --tags "tag1,tag2,tag3" --source "claude-code-session" --category decision --content -
<transcript content>
TRANSCRIPT

6. Emit the one-line save notice described in the Save Notice Convention section. No narration of the dedup or routing steps unless something went wrong.

Important: Do NOT depend on /transcript or any external skill. Generate the transcript yourself from the conversation context you have access to.

---

Import Context Files as Memories

When the user asks to "import my CLAUDE.md", "scan for context files", "import project context", or similar — find context files in the project (CLAUDE.md, AGENTS.md, .cursorrules, .windsurfrules, etc.), split them into logical sections, check for duplicates, confirm with the user, and create memories.

Tag each memory with --source "import:<filename>" for later identification.

See references/import-context.md for the full workflow, supported files list, splitting guidelines, and re-import logic.

---

Memory Templates

When the user asks to "document this feature", "analyze the auth system", "create a memory for the API", or similar — read the relevant source code, select an appropriate template (Feature Guide, Architecture Decision, API Reference, Coding Conventions, Debugging, or Onboarding), draft the memory, check for duplicates, and save with --source "analysis:feature-name".

See references/memory-templates.md for the workflow, template selection table, and structure guidance for each template type.

---

The Core Loop

Each capture follows the same sequence — the detail for every step is in the section named:

1. Notice a capture-worthy moment (When to Capture) — proactively, without being asked.
2. Route to the right workspace (Workspace Routing).
3. Dedup before writing (Dedup-Before-Write Protocol) — default to version create on a strong match; search across workspaces and confirm before any link (Memory Links).
4. Categorize the save (Categories) — always pass --category.
5. Announce in one line (Save Notice Convention) — no narration unless something went wrong.

CLI mechanics — JSON parsing, pagination, stdin piping, error handling, reading large memories in windows — are in references/cli-reference.md.