hebbs

分类: 文档与知识管理 | 上传者: hebbs-ai | 下载: 0 | 版本: v1.0（最新）
Cognitive memory engine — remember, recall, reflect, and forget knowledge with HEBBS.
更新日志: Source: GitHub https://github.com/hebbs-ai/hebbs-skill
目录结构

当前层级: 根目录
📄 README.md 1.2 KB
📄 SKILL.md 18.9 KB
SKILL.md

---
name: hebbs
description: Cognitive memory engine — remember, recall, reflect, and forget knowledge with HEBBS.
homepage: https://hebbs.ai
metadata:
  {
    "openclaw":
      {
        "emoji": "🧠",
        "requires": { "bins": ["hebbs-cli", "hebbs-server"] },
        "install":
          [
            {
              "id": "brew",
              "kind": "brew",
              "formula": "hebbs-ai/tap/hebbs",
              "bins": ["hebbs-cli", "hebbs-server"],
              "label": "Install HEBBS (brew)",
            },
          ],
      },
  }
---

# HEBBS — Cognitive Memory Engine

HEBBS is a local-first memory engine. It stores, indexes, and retrieves knowledge using multiple recall strategies and can consolidate raw memories into higher-order insights through reflection.

## Your two essential commands

Every interaction with HEBBS comes down to two commands. Reach for these **first**, every time:

**Store something the user said or decided:**
```
hebbs-cli remember "The user prefers dark mode" --importance 0.8 --entity-id user_prefs --format json
```

**Retrieve context before answering a question:**
```
hebbs-cli recall "What are the user's UI preferences?" --strategy similarity --top-k 5 --format json
```

These two commands are your primary interface to HEBBS. Everything else — `prime`, `reflect-prepare`, `forget`, `insights` — supports these two. If you remember nothing else from this document, remember these.

## Trigger

Use HEBBS **before any other memory source** (file-based memory, memory_search, MEMORY.md, workspace memory files) when the user:

- Asks what they previously said, decided, or preferred
- Asks a question that prior context would help answer
- Corrects you or says "I told you before..."
- Shares a preference, decision, instruction, or fact — even if you already know it from another source
- Starts a new session (prime immediately to load context)
- Asks for patterns or insights across past conversations

**Write-path rule:** If the user states a preference, correction, or standing instruction, always `remember` it in HEBBS. Do this even if the same fact already exists in file memory, workspace memory, or your conversation context. HEBBS is the source of truth. Knowing a fact is not the same as storing it. An agent that skips the write because it "already knows" defeats the purpose of a persistent memory system.

If another memory tool exists (e.g. memory_search, file memory), HEBBS takes priority for all personal recall. Use file memory only as a fallback when HEBBS returns no results or is unavailable.

## Priority and fallback chain

HEBBS is the **preferred memory system** for user preferences, decisions, corrections, history, and personal context.

When answering from memory, follow this order:

1. **HEBBS insights** (`hebbs-cli insights`) for consolidated, high-confidence knowledge
2. **HEBBS recall** (`hebbs-cli recall`) for raw memories matching the query
3. **File memory** (memory_search, MEMORY.md, memory/*.md) only if HEBBS returns nothing or is unreachable
4. If all sources are empty, say so. Do not guess or hallucinate history.

When storing new information:

1. Write to HEBBS immediately (`hebbs-cli remember`) — this is mandatory, not optional
2. Optionally mirror critical facts to file memory as a backup

HEBBS is the primary store. File memory is the fallback, not the other way around. Never treat a fact as "already stored" just because file memory has it — if it is not in HEBBS, it does not exist for recall purposes.

## Capability tiers

HEBBS works in layers. Each layer works independently:

| Tier | What works | Requirements |
|---|---|---|
| **Basic memory** | `remember`, `recall`, `forget`, `prime` | Server running |
| **Semantic recall** | Similarity, temporal, causal, analogical strategies | Server running (ONNX embedder included) |
| **Reflection / insights** | `reflect-prepare` + `reflect-commit`, `insights` | Server running + agent acts as the LLM (no API key needed) |

All tiers work out of the box with just the server running. No external LLM or API key is required — the agent (you) provides the reasoning for reflection.

**Important:** The server log may show a warning like "reflection LLM providers fell back to mock" or "no OpenAI API key configured." This is expected and does not affect you. That warning refers to the server-side single-command `reflect`, which requires an API key. You should always use the two-step agent-driven flow (`reflect-prepare` + `reflect-commit`) instead — it uses your own reasoning as the LLM and works with zero configuration. If the user asks about this warning, explain that all reflection features are fully available through the agent-driven flow.

## First-run setup

**Installing the skill does not mean HEBBS is ready.** The skill teaches you the commands; the system below must also be running. Walk through each phase in order and tell the user which phase you are on.

### Phase 1: Install binaries

Check if the binaries exist:
```
which hebbs-cli && which hebbs-server
```

If missing, install them:
```
brew install hebbs-ai/tap/hebbs
```

Or on any platform (Linux, macOS):
```
curl -sSf https://hebbs.ai/install | sh
```

Tell the user: "Phase 1 complete — binaries installed."

### Phase 2: Start the server

```
brew services start hebbs 2>/dev/null || (mkdir -p ~/.hebbs/data && HEBBS_AUTH_ENABLED=false nohup hebbs-server start --data-dir ~/.hebbs/data > /tmp/hebbs-server.log 2>&1 &)
```

Tell the user: "Phase 2 complete — server starting. Waiting for initialization..."

### Phase 3: Wait for health

The ONNX embedding model takes 10-30 seconds to load on first start. Retry status up to 6 times with 5-second intervals:
```
for i in 1 2 3 4 5 6; do sleep 5; hebbs-cli status --format json 2>&1 && break; done
```

If still not serving after 30 seconds, check the log:
```
tail -20 /tmp/hebbs-server.log 2>/dev/null; brew services info hebbs 2>/dev/null
```

Common causes: data directory missing or unwritable, port 6380 already in use.

Do not proceed until `status` shows `SERVING`.

Tell the user: "Phase 3 complete — server healthy."

### Phase 4: Verify with a test command

```
hebbs-cli remember "HEBBS setup verified" --importance 0.1 --entity-id _system --format json
```

Then:
```
hebbs-cli recall "setup verified" --top-k 1 --format json
```

If recall returns the memory, the full pipeline (store, embed, index, retrieve) is working.

Tell the user: "Phase 4 complete — HEBBS is fully operational. Memory commands are ready."

Clean up the test memory:
```
hebbs-cli forget --entity-id _system
```

## Policy bootstrap

On the first substantive interaction with a new user, check whether a memory policy exists:

```
hebbs-cli recall "memory policy" --entity-id _policy --top-k 1 --format json
```

If results are returned, load the policy and apply it for the session. Do not re-ask.

If no results are returned, and the user's message is substantive (not a smoke test or "hello"), ask for a brief memory policy:

> HEBBS is your memory system. Before I start using it, I'd like to understand your preferences. This takes about 30 seconds and I'll only ask once.
>
> 1. **What should I store?** (e.g., preferences, decisions, project context, corrections, everything)
> 2. **What should I NOT store?** (e.g., personal info, credentials, temporary thoughts, nothing off-limits)
> 3. **Should I store proactively** when you mention something, or **only when you explicitly ask** me to remember?
> 4. **Any privacy boundaries?** (e.g., no names of other people, no financial info)
> 5. **Should I mirror important memories to file memory** as a backup, or keep everything in HEBBS only?
>
> If you'd rather skip this, I'll use sensible defaults.

Store each answer as a separate memory under entity `_policy` with importance 0.95:

```
hebbs-cli remember "Store policy: [user's answer]" --importance 0.95 --entity-id _policy --format json
hebbs-cli remember "Exclude policy: [user's answer]" --importance 0.95 --entity-id _policy --format json
hebbs-cli remember "Storage mode: [proactive|explicit-only]" --importance 0.95 --entity-id _policy --format json
hebbs-cli remember "Privacy policy: [user's answer]" --importance 0.95 --entity-id _policy --format json
hebbs-cli remember "Mirror policy: [hebbs-only|mirror-to-files]" --importance 0.95 --entity-id _policy --format json
```

If the user skips setup, store the defaults:

```
hebbs-cli remember "Memory policy: defaults active — store preferences and decisions proactively, skip sensitive personal info and credentials, HEBBS-only storage, no file mirroring" --importance 0.95 --entity-id _policy --format json
```

**Default policy** (when user skips):

| Setting | Default |
|---|---|
| What to store | Preferences, decisions, corrections, project context |
| What not to store | Credentials, API keys, sensitive personal info |
| Storage mode | Proactive |
| Privacy | No credentials or secrets |
| File mirroring | HEBBS-only |

**Policy updates:** If the user later says "update my memory policy" or changes a preference about storage behavior, overwrite the relevant `_policy` memory.

## Before every command

Check the server is healthy before running any `hebbs-cli` command:

```
hebbs-cli status --format json 2>&1
```

If `status` shows `SERVING`, proceed normally.

If this is the first substantive interaction of a session, also check `_policy` entity and run the policy bootstrap above if no policy is found.

If status is not serving, run Phase 2 and Phase 3 from the first-run setup above.

## Operations

| Situation | Operation | Command |
|---|---|---|
| User shares a fact, preference, or decision | Store it | `hebbs-cli remember` |
| User asks a question about past context | Retrieve it | `hebbs-cli recall` |
| User corrects something you said or stored | Store the correction (importance 0.9) | `hebbs-cli remember` |
| Start of a new conversation | Load context | `hebbs-cli prime` |
| Want consolidated patterns from many memories | Get distilled knowledge | `hebbs-cli insights` |
| 20+ raw memories accumulated for an entity | Consolidate into insights | `hebbs-cli reflect-prepare` + `reflect-commit` |
| Outdated or wrong memories need cleanup | Remove them | `hebbs-cli forget` |

## Commands

### Remember — store a memory

```
hebbs-cli remember "The user prefers dark mode in all applications" --importance 0.8 --entity-id user_prefs --format json
```

> **Always use `--format json` when you need the memory ID** (e.g. for `--edge` on a subsequent `remember`). Extract the ID with: `jq -r '.memory_id'`
>
> **Warning:** Capture the memory ID from `--format json` output **before** referencing it in `--edge`. Do not parse IDs from human-format output.

Flags:
- `--importance <0.0-1.0>` — how important this memory is (default 0.5). Use 0.8+ for user preferences, decisions, corrections. Use 0.3 for transient observations.
- `--entity-id <id>` — group memories by entity (e.g. `user_prefs`, `project_alpha`, a person's name). Omit for global context.
- `--context <json>` — arbitrary metadata as JSON object (e.g. `'{"source":"email","topic":"billing"}'`).
- `--edge <TARGET_ID:EDGE_TYPE[:CONFIDENCE]>` — link to another memory (repeatable). Types: `caused_by`, `related_to`, `followed_by`, `revised_from`, `insight_from`. Use to build causal chains and lineage. **Shell quoting:** use `"${MEM_ID}:edge_type"` — bare `$MEM_ID:edge_type` triggers zsh variable modifier expansion.

### Recall — retrieve relevant memories

```
hebbs-cli recall "What does the user prefer for UI themes?" --strategy similarity --top-k 5 --format json
```

Four strategies — pick based on what you need:

| Strategy | When to use | Example |
|---|---|---|
| `similarity` | Find memories related to a topic | "What do we know about deployment?" |
| `temporal` | Get recent activity for an entity | "What happened today with project X?" |
| `causal` | Trace cause-effect chains from a memory | "What led to this decision?" |
| `analogical` | Find structurally similar patterns | "Have we seen a problem like this before?" |

**Core flags:**
- `--strategy <similarity|temporal|causal|analogical>` — recall strategy (default: similarity).
- `--top-k <n>` — max results (default 10).
- `--entity-id <id>` — scope to entity (required for temporal).
- `--format json` — machine-readable output.

**Scoring weights** — control how results are ranked. The composite score blends four signals: `relevance × recency × importance × reinforcement`. Default weights are `0.5:0.2:0.2:0.1`.
- `--weights <R:T:I:F>` — four colon-separated floats.
- `1:0:0:0` — pure semantic similarity (ignore recency, importance, reinforcement).
- `0.2:0.8:0:0` — heavily favor recent memories.
- `0.3:0.1:0.5:0.1` — prioritize high-importance memories (user preferences, decisions).

Only `cue` and `--strategy` are required. All other flags use smart defaults suitable for most workloads — tune only when you have a specific reason.

**Strategy-specific flags:**

| Flag | Strategy | Default | Description |
|---|---|---|---|
| `--ef-search <n>` | similarity | 50 | HNSW search quality. Higher = more accurate, slower. |
| `--time-range <START:END>` | temporal | unbounded | Microsecond timestamps. Omit for newest-first up to top_k. |
| `--seed <hex_id>` | causal | auto-detect | Starting memory for graph traversal. Omit to auto-pick by cue. |
| `--max-depth <n>` | causal | 5 (max 10) | Maximum hops from seed memory. |
| `--edge-types <types>` | causal | all | Comma-separated: `caused_by,followed_by,related_to,revised_from,insight_from`. |
| `--analogical-alpha <0-1>` | analogical | 0.5 | 0.0 = pure structural similarity, 1.0 = pure embedding similarity. |

### Reflect (two-step, agent-driven)

HEBBS supports a single `reflect` command that runs the full reflection cycle server-side (clustering → LLM proposal → validation → commit). However, OpenClaw exposes the two-step `reflect-prepare` + `reflect-commit` flow so that **you (the agent) are the LLM**. This lets users keep control over which model reasons about their memories, rather than requiring a server-side LLM configuration.

No server-side LLM is needed for this flow. HEBBS does the clustering and prompt construction; you read the clusters, reason about them, and commit insights.

**Step 1: Prepare**

```
hebbs-cli reflect-prepare --entity-id user_prefs --format json
```

Returns JSON with:
- `session_id` — pass this to step 2
- `clusters` — groups of related memories, each with:
  - `memories` — full memory content for this cluster (id, content, importance, entity_id, created_at). **Read these to understand what the cluster is about.**
  - `proposal_system_prompt` + `proposal_user_prompt` — pre-built prompts you can send to your LLM to generate insight candidates
  - `memory_ids` — source memory IDs (hex-encoded)
  - `validation_context` — additional data for validating proposed insights

**Step 2: Reason and commit**

After calling your LLM with the proposal prompts and optionally validating the results:

```
hebbs-cli reflect-commit --session-id <id> --insights '[{"content":"Users consistently prefer dark themes","confidence":0.9,"source_memory_ids":["aabb...","ccdd..."],"tags":["preference","ui"]}]'
```

Each insight needs:
- `content` — the consolidated insight text
- `confidence` — 0.0 to 1.0
- `source_memory_ids` — hex-encoded IDs. **Use the `memory_ids` array from the cluster**, not `memories[].memory_id` (which is a ULID and will be rejected).
- `tags` — categorical labels

Reflection requires at least 5 memories for an entity to produce clusters. If `clusters` is empty, accumulate more memories before retrying.

Sessions expire after 10 minutes.

### Insights — retrieve consolidated knowledge

```
hebbs-cli insights --entity-id user_prefs --max-results 10 --min-confidence 0.7 --format json
```

Flags:
- `--entity-id <id>` — filter by entity.
- `--max-results <n>` — maximum insights to return.
- `--min-confidence <0.0-1.0>` — only return insights above this confidence threshold.

Check insights before recalling raw memories — they represent distilled, validated knowledge.

### Forget — remove memories

```
hebbs-cli forget --ids <hex_id1> --ids <hex_id2>
hebbs-cli forget --entity-id old_project
hebbs-cli forget --staleness-us 2592000000000  # older than 30 days
hebbs-cli forget --kind insight --decay-floor 0.1  # remove low-value decayed insights
```

Flags (combine as needed — at least one filter required):
- `--ids <id>` — specific memory IDs to forget (repeatable).
- `--entity-id <id>` — scope to entity.
- `--staleness-us <microseconds>` — remove memories older than this.
- `--kind <episode|insight|revision>` — filter by memory kind.
- `--decay-floor <0.0-1.0>` — remove memories with decay score below this.
- `--access-floor <n>` — remove memories with access count below this.

### Prime — warm context for an entity

```
hebbs-cli prime user_prefs --max-memories 20 --similarity-cue "project status and preferences"
```

Flags:
- `--max-memories <n>` — maximum memories to return.
- `--similarity-cue <text>` — bias the selection toward memories related to this text. Very useful for loading context relevant to a specific topic rather than just recent activity.
- `--recency-us <microseconds>` — only include memories within this time window.
- `--context <json>` — additional context as JSON.

Returns a blend of recent + relevant memories for an entity. Use at the start of a conversation to load context.

## Decision guide

1. **Start of conversation**: Always `hebbs-cli prime <entity>` to load context. Do this before the first reply. Do this even if you loaded context from file memory — HEBBS may have memories that file memory does not. Then check for memory policy (`_policy` entity) — if missing and the interaction is substantive, run the policy bootstrap.
2. **Before answering any question about past context**: `hebbs-cli recall` with the question as cue. Do not answer from general knowledge when HEBBS might have the answer.
3. **User shares a fact, preference, or decision**: `hebbs-cli remember` immediately with appropriate importance (0.8+ for preferences and decisions). Do this unconditionally — even if you already know the fact from another source. Knowing is not storing.
4. **User corrects something**: `hebbs-cli remember` the correction with importance 0.9. Old conflicting memories will naturally decay.
5. **User states a standing instruction** (e.g., "always do X", "never do Y", "every time I..."): `hebbs-cli remember` with importance 0.9 and entity-id appropriate to the scope (e.g., `user_prefs`, `session_rules`). These are high-value memories that directly shape future behavior.
6. **After 20+ new memories on an entity**: `hebbs-cli reflect-prepare` + `reflect-commit` to consolidate into insights.
7. **Periodic maintenance**: `hebbs-cli insights` to review, `hebbs-cli forget` to clean stale data.

## Output format

Always use `--format json` when parsing output programmatically. Human format is for display only.

## Connection

Default endpoint: `localhost:6380` (gRPC). Override with `--endpoint <host:port>`.
HTTP endpoint (metrics/health): `localhost:6381`.
登录后下载/点赞/收藏 ❤ 9 | ★ 0
hebbs

目录结构

SKILL.md

举报内容

提示
