memory

package
v0.52.1 Latest Latest
Warning

This package is not in the latest version of its module.

 Go to latest Published: May 25, 2026 License: MIT Imports: 13 Imported by: 0

Documentation

Overview

Package memory provides persistent, agent-managed memory across sessions.

Architecture

Three tiers:

  1. Facts — Two typed files (user.md, env.md) with character caps, injected into the system prompt as a frozen snapshot. Managed by the agent via the memory tool (add/replace/remove/consolidate/read).

  2. Buffer — In-memory ring buffer on the Session struct. One-line summaries appended after each turn. Injected only when non-empty.

  3. Episodes — LLM-extracted durable facts written after sessions with ≥3 turns. Searchable via memory(search=...).

Merge-on-Write

When adding a fact, go-vector's RandomProjections provides a fast similarity check vs existing entries. cos > 0.7 = auto-merge, cos < 0.3 = auto-add, 0.3-0.7 = SimpleCall judgment. This saves ~80% of LLM calls on writes.

Index

Constants

View Source 
const (
	// MergeThreshold is the cosine similarity threshold above which entries
	// are considered duplicates and auto-merged.
	MergeThreshold float32 = 0.7

	// AddThreshold is the cosine similarity below which entries are
	// considered distinct and auto-added without LLM judgment.
	AddThreshold float32 = 0.3
)

MergeThresholds for merge-on-write classification.

Variables

This section is empty.

Functions

func BoolPtr 

func BoolPtr(b bool) *bool

BoolPtr returns a pointer to a bool value.

func FormatBufferLine 

func FormatBufferLine(role, message string) string

FormatBufferLine creates a timestamped buffer line. Format: "HH:MM role message"

func ScanContent 

func ScanContent(content string) error

ScanContent checks memory content for security threats. Returns an error if the content contains patterns that could compromise the agent.

Checks:

  • Invisible Unicode characters (zero-width spaces, direction overrides, BOM)
  • Prompt injection markers ("ignore previous instructions", etc.)
  • Credential exfiltration patterns (API keys, private keys, bearer tokens)

Types

type Buffer 

type Buffer struct {
	// contains filtered or unexported fields
}

Buffer is a simple ring buffer for session-level turn summaries. Thread-safe only if accessed from a single goroutine (the loop engine owns the buffer and serializes all access).

func NewBuffer 

func NewBuffer(cap int) *Buffer

NewBuffer creates a ring buffer with the given capacity. cap=0 means disabled (all appends are silently discarded).

func (*Buffer) Append 

func (b *Buffer) Append(line string)

Append adds a line to the buffer. If the buffer is at capacity, the oldest line is evicted first.

func (*Buffer) Cap 

func (b *Buffer) Cap() int

Cap returns the maximum number of lines.

func (*Buffer) Clear 

func (b *Buffer) Clear()

Clear removes all lines from the buffer.

func (*Buffer) Len 

func (b *Buffer) Len() int

Len returns the current number of lines.

func (*Buffer) Lines 

func (b *Buffer) Lines() []string

Lines returns a copy of the current buffer contents (oldest first).

type EpisodeMeta 

type EpisodeMeta struct {
	SessionID string    `json:"session_id"`
	Turns     int       `json:"turns"`
	CreatedAt time.Time `json:"created_at"`
	Summary   string    `json:"summary"` // truncated for index listing
}

EpisodeMeta holds metadata for a single episode.

type EpisodeStore 

type EpisodeStore struct {
	// contains filtered or unexported fields
}

EpisodeStore manages on-disk episode summaries (Tier 3 memory). Written after sessions with sufficient turns, searchable via SimpleCall. Index operations are protected by a mutex to prevent TOCTOU races between concurrent sessions sharing the same memory directory.

The in-memory idxCache avoids reading + unmarshalling index.json from disk on every FormatEpisodeContext call (which fires every loop turn). The cache is invalidated after every write.

func NewEpisodeStore 

func NewEpisodeStore(dir string, rankFn RankStrategy) *EpisodeStore

NewEpisodeStore creates an EpisodeStore rooted at dir. If rankFn is nil, a default ranker is used (SimpleCall-based — requires LLM client).

func (*EpisodeStore) Read 

func (e *EpisodeStore) Read(sessionID string) (string, error)

Read returns the full summary content for a session. sessionID is validated for path traversal before use.

func (*EpisodeStore) ReadIndex 

func (e *EpisodeStore) ReadIndex() ([]EpisodeMeta, error)

ReadIndex reads the episode index from disk. Returns empty slice if the index file doesn't exist yet. Entries are ordered newest-first.

Uses an in-memory cache to avoid disk I/O on every Search() call. The cache is invalidated after every writeIndex().

func (*EpisodeStore) Search 

func (e *EpisodeStore) Search(query string, limit int) ([]EpisodeMeta, error)

Search returns the most relevant episodes for a query, ranked by the configured RankStrategy. Limited to limit results.

func (*EpisodeStore) Write 

func (e *EpisodeStore) Write(sessionID, summary string, turns int) error

Write stores an episode summary for a session. Creates the episodes directory and updates the index. sessionID is validated for path traversal before use.

func (*EpisodeStore) WriteIfEnough 

func (e *EpisodeStore) WriteIfEnough(sessionID, summary string, turns int) error

WriteIfEnough calls Write only if turns >= threshold. Returns nil without writing if the threshold isn't met.

type FactStore 

type FactStore struct {
	// contains filtered or unexported fields
}

FactStore manages typed fact files (user.md and env.md) with character caps, duplicate prevention, and entry-level CRUD via substring matching. All write operations are protected by a mutex to prevent TOCTOU races between concurrent sessions sharing the same memory directory. The mutex guards only the in-memory read+parse+modify phase; the final disk write happens outside the lock to avoid blocking other sessions during file I/O.

func NewFactStore 

func NewFactStore(dir string, capUser, capEnv int) *FactStore

NewFactStore creates a FactStore rooted at dir. Fact files are stored as dir/user.md and dir/env.md. Caps limit total file size.

func (*FactStore) Add 

func (f *FactStore) Add(target, content string) error

Add appends a new entry to a fact file. Returns error if:

  • target is invalid
  • content is empty
  • content already exists (dedup)
  • adding would exceed the character cap

func (*FactStore) Entries 

func (f *FactStore) Entries(target string) ([]string, error)

Entries returns the individual entries as a string slice.

func (*FactStore) Read 

func (f *FactStore) Read(target string) (string, error)

Read returns the full content of a fact file. Returns empty string if the file doesn't exist yet.

func (*FactStore) Remove 

func (f *FactStore) Remove(target, oldText string) error

Remove finds an entry by substring match and removes it. Returns error if the substring doesn't match exactly one entry.

func (*FactStore) Replace 

func (f *FactStore) Replace(target, oldText, content string) error

Replace finds an entry by substring match and replaces it with new content. Returns error if the substring doesn't match exactly one entry.

type LLMClient 

type LLMClient interface {
	SimpleCall(ctx context.Context, system, user string) (string, error)
}

LLMClient abstracts the LLM calls needed by the memory system (SimpleCall for consolidation, episode extraction, and search).

type MemoryConfig 

type MemoryConfig struct {
	Enabled               *bool   `json:"enabled,omitempty"`
	FactsLimitUser        int     `json:"facts_limit_user,omitempty"`
	FactsLimitEnv         int     `json:"facts_limit_env,omitempty"`
	BufferLines           int     `json:"buffer_lines,omitempty"`
	BufferEnabled         *bool   `json:"buffer_enabled,omitempty"`
	MergeOnWrite          *bool   `json:"merge_on_write,omitempty"`
	ExtractOnEnd          *bool   `json:"extract_on_end,omitempty"`
	LLMSearch             *bool   `json:"llm_search,omitempty"`
	LLMExtract            *bool   `json:"llm_extract,omitempty"`
	LLMConsolidate        *bool   `json:"llm_consolidate,omitempty"`
	MergeThreshold        float32 `json:"merge_threshold,omitempty"`
	AddThreshold          float32 `json:"add_threshold,omitempty"`
	MinTurnsForExtraction int     `json:"min_turns_for_extraction,omitempty"`
}

MemoryConfig holds configuration for the memory system. Mirrors the JSON config section. Bool fields use *bool so that JSON omitempty can distinguish "not set" (nil) from "explicitly false" (pointer to false).

func DefaultMemoryConfig 

func DefaultMemoryConfig() MemoryConfig

DefaultMemoryConfig returns sensible defaults.

type MemoryManager 

type MemoryManager struct {
	// contains filtered or unexported fields
}

MemoryManager orchestrates all three tiers of memory: Facts (durable, in system prompt), Buffer (session-level context), and Episodes (on-disk extracted summaries with search).

func NewMemoryManager 

func NewMemoryManager(memoryDir string, llc LLMClient, cfg MemoryConfig) *MemoryManager

NewMemoryManager creates a fully wired MemoryManager. If llc is nil, LLM-dependent features (consolidation, episode search) degrade gracefully (no LLM call, fallback behavior).

func (*MemoryManager) AddFact 

func (m *MemoryManager) AddFact(target, content string) error

AddFact appends a new fact entry. Performs:

  1. Security scan (reject if dangerous)
  2. Dedup (FactStore handles this)
  3. Merge-on-write using go-vector RP if MergeOnWrite is enabled
  4. Fits the merge detector after mutation (incrementally — only re-embeds the changed entry instead of all entries, avoiding a double disk-read).

func (*MemoryManager) AppendBuffer 

func (m *MemoryManager) AppendBuffer(role, message string)

AppendBuffer adds a turn summary to the in-memory ring buffer.

func (*MemoryManager) BuildSystemPrompt 

func (m *MemoryManager) BuildSystemPrompt() string

BuildSystemPrompt returns the memory section to inject into the system prompt. Returns empty string if memory is disabled or nothing to show.

Security: all content is scanned for injection patterns. If detected, the section is prefixed with an explicit warning that the content is data, not instructions. Even clean content is wrapped with anti-injection framing.

func (*MemoryManager) ClearBuffer 

func (m *MemoryManager) ClearBuffer()

ClearBuffer resets the buffer for a new session.

func (*MemoryManager) Consolidate 

func (m *MemoryManager) Consolidate(target string) error

Consolidate uses the LLM to merge related entries in a target file for better density. Falls back to no-op if LLM is unavailable or LLMConsolidate is disabled in config.

func (*MemoryManager) FormatEpisodeContext 

func (m *MemoryManager) FormatEpisodeContext(query string) string

FormatEpisodeContext searches episodes with a recency-based ranker (no LLM — safe for per-turn use without recursion risk) and returns formatted context to inject as a system message. Returns empty string if no episodes found or memory is disabled.

func (*MemoryManager) GetBuffer 

func (m *MemoryManager) GetBuffer() []string

GetBuffer returns the current buffer lines (for system prompt injection).

func (*MemoryManager) OnSessionEnd 

func (m *MemoryManager) OnSessionEnd(sessionID string, turns int, messages []string)

OnSessionEnd is called when a session ends. If turns >= threshold, extracts a narrative session summary using the LLM and stores it as an episode. sessionID is validated for path traversal before any file I/O.

func (*MemoryManager) ReadFacts 

func (m *MemoryManager) ReadFacts() (userContent, envContent string, err error)

ReadFacts returns the full content of user and env fact files.

func (*MemoryManager) RemoveFact 

func (m *MemoryManager) RemoveFact(target, oldText string) error

RemoveFact removes a fact entry by substring.

func (*MemoryManager) ReplaceFact 

func (m *MemoryManager) ReplaceFact(target, oldText, content string) error

ReplaceFact replaces an existing fact entry.

func (*MemoryManager) RestoreBuffer 

func (m *MemoryManager) RestoreBuffer(lines []string)

RestoreBuffer loads buffer lines from a saved slice (e.g., from session).

func (*MemoryManager) SearchEpisodes 

func (m *MemoryManager) SearchEpisodes(query string, limit int) ([]EpisodeMeta, error)

SearchEpisodes returns the most relevant episodes for a query.

type MemoryTool 

type MemoryTool struct {
	// contains filtered or unexported fields
}

MemoryTool wraps a MemoryManager as a odek-compatible Tool.

func NewMemoryTool 

func NewMemoryTool(mm *MemoryManager) *MemoryTool

NewMemoryTool creates a tool that exposes memory CRUD + search to the agent.

func (*MemoryTool) Call 

func (t *MemoryTool) Call(args string) (string, error)

func (*MemoryTool) Description 

func (t *MemoryTool) Description() string

func (*MemoryTool) Name 

func (t *MemoryTool) Name() string

func (*MemoryTool) Schema 

func (t *MemoryTool) Schema() any

type MergeDetector 

type MergeDetector struct {
	// contains filtered or unexported fields
}

MergeDetector uses RandomProjections to quickly estimate whether a new fact entry overlaps with existing entries. This avoids ~80% of LLM calls during merge-on-write.

Lifecycle:

  1. NewMergeDetector(dims) — creates RP embedder
  2. Fit(corpus) — builds vocabulary from existing entries
  3. Classify(entry) → action + similarIdx + similarity
  4. After facts change → Fit(newCorpus) to rebuild vocabulary

Thresholds control the classification:

  • mergeThreshold: cosine above this → auto-merge (default 0.7)
  • addThreshold: cosine below this → auto-add (default 0.3)
  • Between thresholds: "judge" — requires LLM to decide

func NewMergeDetector 

func NewMergeDetector(dims int) *MergeDetector

NewMergeDetector creates a MergeDetector with the given output dimensionality for the RP embedder. Pass 0 for default (256). Uses default thresholds (0.7 merge, 0.3 add).

func NewMergeDetectorWithThresholds 

func NewMergeDetectorWithThresholds(dims int, mergeThreshold, addThreshold float32) *MergeDetector

NewMergeDetectorWithThresholds creates a MergeDetector with custom thresholds.

func (*MergeDetector) AppendEntry 

func (m *MergeDetector) AppendEntry(entry string)

AppendEntry adds a single entry to the corpus. Only the new entry is embedded, avoiding a full re-embed of all existing entries. The RP vocabulary is still refreshed so new tokens from the entry are available for future Classify calls.

func (*MergeDetector) Classify 

func (m *MergeDetector) Classify(entry string) (action string, similarIdx int, similarity float32)

Classify returns the merge decision for a new entry vs the fitted corpus.

Returns:

  • action: "merge" | "add" | "judge" | "nobody"
  • similarIdx: index of the most similar corpus entry (for merge/judge)
  • similarity: cosine similarity [0, 1]

"nobody" means the corpus is empty — there's nothing to compare against.

func (*MergeDetector) Corpus 

func (m *MergeDetector) Corpus() []string

Corpus returns the current corpus (for inspection).

func (*MergeDetector) Fit 

func (m *MergeDetector) Fit(corpus []string)

Fit builds the RP vocabulary and pre-computes embeddings for all corpus entries. Call whenever facts change (after add/replace/remove).

func (*MergeDetector) ReplaceEntry 

func (m *MergeDetector) ReplaceEntry(idx int, entry string)

ReplaceEntry replaces an entry at the given index. Only the changed entry is re-embedded, avoiding a full re-embed of all existing entries.

type RankStrategy 

type RankStrategy func(query string, episodes []EpisodeMeta) ([]EpisodeMeta, error)

RankStrategy is an injectable function for ranking episodes by relevance to a query. The default implementation uses SimpleCall; tests can inject a deterministic mock.

func NewLLMRanker 

func NewLLMRanker(llm LLMClient) RankStrategy

NewLLMRanker creates a RankStrategy that uses an LLM client to rank episodes by semantic relevance to the query. Falls back to recency ordering if the LLM call fails or returns unparseable output.

func NewRPRanker 

func NewRPRanker(dims int) RankStrategy

NewRPRanker creates a RankStrategy that uses RandomProjections (go-vector) for semantic similarity search over episodes. No LLM calls — pure vector math. Falls back to recency ordering if RP fitting fails.

The ranker fits the RP embedder on episode summaries on each call. With the typical number of episodes (< 100, 120-char summaries each), fitting takes < 1ms and is negligible compared to LLM latency.

Source Files

View all Source files

Jump to

Keyboard shortcuts

? : This menu
/ : Search site
f or F : Jump to
y or Y : Canonical URL
go.dev uses cookies from Google to deliver and enhance the quality of its services and to analyze traffic. Learn more.