Memory

All storage lives under your Flowly home (~/.flowly by default, or ~/.flowly/profiles/<name>/ for named profiles). Configuration is at ~/.flowly/config.json under agents.defaults.memorySearch and agents.defaults.memoryNudgeInterval.

The three layers

1. Raw markdown memory

Human- and agent-readable .md files on disk:

These are plain markdown you can open and edit directly. MEMORY.md is the curated long-term store; daily notes capture per-day context.

2. Memory search index

The same markdown files chunked into a SQLite index for fast retrieval:

• Path: ~/.flowly/memory_index.sqlite (plus -wal / -shm sidecar files).
• Files are chunked (default 400 tokens per chunk, 80-token overlap), with each chunk stored alongside an FTS5 keyword entry and, when an embedding provider is active, a JSON-stored embedding vector.
• Indexing is lazy and debounced to 1 second — it runs on each memory_search call. There is no always-on file watcher, so edits to your markdown are picked up on the next search.
• Re-indexing is change-detected via a SHA-256 hash per file; only changed files are re-chunked.

3. Knowledge graph

A structured temporal triple store of facts about people, companies, projects, and relationships, at ~/.flowly/knowledge_graph.sqlite3. See knowledge-graph.md for the full data model and tool.

How memory enters the system prompt

When memory is not skipped, the agent's context is built with:

• MEMORY.md is always injected as a # Memory section — no tool call required. It is scanned for prompt-injection payloads before injection.
• Knowledge graph summary is injected as # Knowledge Graph (top entities by current-fact count). See the knowledge-graph doc for the workspace caveat.
• Recent daily notes (last 3 days) are injected as # Recent Notes only when the memory_search tool is not available. When memory_search is enabled, the agent is expected to search on demand instead.

Automatic curation (background self-review)

Flowly curates memory automatically. Every memoryNudgeInterval user turns (default 10; set to 0 to disable), a fire-and-forget background "self-review" subagent runs. It receives the last 20 messages plus the current MEMORY.md (truncated to ~8000 chars) and the knowledge-graph summary, and is instructed to:

• Extract structured facts into the knowledge_graph tool.
• Append only genuinely-new free-form preferences via memory_append.
• Otherwise reply "Nothing to save."

The review runs silently and does not interrupt your session. This is the primary automatic write path; the main agent may also write via tools directly.

Tools

memory_search

Searches the memory index. Always call before answering about prior conversations, preferences, names, dates, or projects.

Returns JSON:

{
  "results": [
    { "path": "...", "lines": "...", "score": 0.0, "snippet": "..." }
  ],
  "provider": "openai",
  "vector_enabled": true
}

vector_enabled tells you whether semantic vector scoring actually ran for this query (see the embeddings caveat below).

memory_get

Reads exact lines from a memory file (falls back to the index if not on disk).

memory_append

The wired markdown writer — appends to MEMORY.md (not daily notes). It enforces several guards:

• Content-injection scanning before writing.
• Exact-duplicate rejection (normalized SHA-256) and near-duplicate rejection (character-trigram Jaccard similarity ≥ 0.75).
• Size cap of 12000 chars; oldest timestamped entries are evicted first.
• Each entry is delimited by an HTML-comment timestamp (<!-- YYYY-MM-DD ... -->).

For structured facts (people, companies, relationships), the agent is directed to use the knowledge_graph tool instead of memory_append.

How search works

memory_search always runs FTS5 BM25 keyword search. When an OpenAI embedding provider is active, it also runs vector (cosine) similarity over chunk embeddings and merges the two with a hybrid score:

final = vectorWeight * vector_score + textWeight * text_score

Results below minScore are dropped; the rest are sorted descending and capped to maxResults. Vector search is a brute-force linear scan over all chunk embeddings (no ANN index) — fine for typical memory sizes, but it scales linearly with history.

Embeddings caveat — read this

Search behavior depends on which embedding provider is configured:

• OpenAI embeddings configured → true hybrid BM25 + vector search (text-embedding-3-small, 1536-dim).
• Gemini selected → embeddings are non-functional after the litellm migration. Selecting Gemini logs a warning and silently falls back to FTS5 keyword-only search.
• No provider / provider: none → FTS5 keyword-only search.
• provider: auto picks OpenAI then Gemini from your configured keys. If only a Gemini key is present, auto resolves to Gemini and you get keyword-only search despite a provider appearing "configured."

To verify what actually ran, check the provider and vector_enabled fields in memory_search output.

Configuration

Under agents.defaults.memorySearch in ~/.flowly/config.json (keys are camelCase on disk):

agents.defaults.memoryNudgeInterval (default 10) controls the background self-review cadence; 0 disables it.

Session search

Distinct from memory, the session_search tool searches past conversation transcripts in ~/.flowly/session_index.sqlite (FTS5 over messages). It operates in three modes inferred from its arguments:

• DISCOVER — pass query: keyword search across sessions, deduped by session, returning a snippet plus surrounding context and an anchor_id.
• SCROLL — pass target_session + around_message_id (optional window, default 5, max 20): a window of messages around an anchor. Refuses to scroll the active session.
• BROWSE — no args: recent sessions chronologically.

None of these are required. The runtime injects the active conversation id internally so the agent cannot scroll into its own in-context session.

Governance, the dreamer & the flowly memory CLI

Behind the markdown, memories are governed in a SQLite store (~/.flowly/memory_governance.sqlite3) with a real lifecycle — candidates, calibrated trust scores, conflict reconciliation, and an append-only audit trail. MEMORY.md is the human-readable projection of the active set. Three processes write into it, each with a distinct job — don't confuse them:

Per-turn self-review (live, single-session)

The automatic curation above: every memoryNudgeInterval turns (default 10) a background subagent reads the recent messages and appends genuinely-new facts via memory_append / knowledge_graph. Fast and greedy — it captures facts you state explicitly the moment you say them, straight to active.

Cross-session "dreaming" (offline, across sessions)

A separate pass that reads conversation deltas across sessions (watermarked, so it never re-reads the same messages), extracts durable candidates the live path missed, and reconciles them against what's already known — both the governed store and your USER.md profile — before committing. On by default (agents.defaults.memoryDreaming.enabled = true), it fires:

• after 30 min with no user activity (idle — background heartbeats don't count),
• once daily at 03:30 local,
• every 10 user turns (a coarse pass),
• or on demand: flowly memory dream, or the desktop / iOS "Learn from chats" button.

Each candidate is routed by confidence and privacy:

• explicitly stated + high confidence → active (remembered immediately);
• an explicit contradiction of a known fact → supersedes it (the old one is closed and linked to the new active one);
• inferred, mid-confidence, or sensitive → the review queue (needs_review), for you to approve;
• already known (same fact in the store or USER.md) → skipped, never duplicated.

The review queue surfaces in the CLI (flowly memory review), the desktop memory panel, and the iOS app (a sheet on app open). Accepting promotes an item to active; rejecting discards it.

Tune it under agents.defaults.memoryDreaming (camelCase on disk): enabled, commitMode ("selective" default — "manual" queues everything for review, "aggressive" lowers the bar), idleMinutes (30), dailyEnabled (true) + dailyTime ("03:30"), turnInterval (10), autoFloor (0.80 — at/above ⇒ auto-active when unconflicted and not sensitive), reviewFloor (0.55 — below ⇒ dropped), maxMessagesPerRun (500). Set turnInterval: 0 to drop the per-turn pass and rely on idle + daily only; enabled: false turns the layer off entirely. The full config table is in Configuration.

Consolidation (cleanup of the existing store)

Distinct from dreaming, consolidation is an LLM-driven cleanup of what's already stored — it merges cross-key duplicate facts, retires free-form that duplicates the knowledge graph, and marks outdated notes stale. It runs in the background (every ~50 turns / ~30 min, gated on new writes since the last pass) and on demand via flowly memory consolidate or the desktop "Clean now" button. Configured under the same block: autoConsolidate, consolidateTurnInterval (50), consolidateEveryMinutes (30).

The CLI

Inspect and correct everything from the CLI (or /memory in a chat):

flowly memory list                  # active long-term memories
flowly memory review                # the review queue (pending candidates)
flowly memory accept <id>           # or: reject <id>
flowly memory dream                 # learn from recent chats now (cross-session)
flowly memory feedback <id>         # 👍/👎 to retune a memory's trust score
flowly memory correct <id> "..."    # fix a memory's content
flowly memory undo                  # revert the last change
flowly memory consolidate           # merge duplicates / retire stale now
flowly memory refresh               # rebuild the MEMORY.md block from the store
flowly memory status                # store statistics

See CLI commands for the full group.

Related

• Knowledge graph
• Delegation
• Features overview
• Sessions
• Tools reference