GitHub - hebbs-ai/hebbs

4 recall strategies · Native consolidation · Automatic decay · One skill file, zero config

brew install hebbs-ai/tap/hebbs

curl -sSf https://hebbs.ai/install | sh

Without HEBBS:                          With HEBBS:

1. Choose a vector DB                   brew install hebbs-ai/tap/hebbs
2. Set up embedding pipeline
3. Write storage layer                  Done.
4. Write retrieval layer
5. Add temporal logic                   Your agent now has:
6. Add graph traversal                  - 4 recall strategies
7. Wire it all together                 - Temporal + causal + analogical
8. Handle decay manually                - Native decay & reinforcement
9. Build consolidation pipeline         - Automatic consolidation
10. Maintain 4 services                 - Works with Claude & OpenClaw

~2,000 lines of glue                    0 lines of glue

Raw memories (episodes):
  "Customer asked about pricing"
  "Customer mentioned competitor X"
  "Customer objected to annual commitment"
  "Deal lost to competitor X"

          | reflect (automatic consolidation)

Insight (with lineage):
  "Deals mentioning competitor X with pricing objections
   have 73% loss rate when annual commitment is pushed early"
   [confidence: 0.84, sources: 4 memories, tags: sales, pricing]

Session 1: Agent runs eval queries against the vault
           Baseline: 54% keyword recall (default settings)

           Agent analyzes failures:
           - k too low → increase to 10
           - cue too generic → expand with entity names
           - wrong strategy → switch to temporal for timeline queries

           After tuning: 84% keyword recall

           Agent stores what worked:
           hebbs remember "RETRIEVAL-INSTRUCTION: For compliance queries,
           include entity names and expand acronyms. Use k=10."
           --importance 0.9 --entity-id retrieval-instructions

Session 2: Agent loads stored strategies at conversation start
           Applies learned tuning automatically
           Retrieval is better from the first query

# Initialize with OpenAI (recommended)
hebbs init . --provider openai --key $OPENAI_API_KEY

# Or with other providers
hebbs init . --provider anthropic --key $ANTHROPIC_API_KEY
hebbs init . --provider ollama

hebbs index .                         # index your markdown files
hebbs recall "your question here"     # recall with any strategy
hebbs panel                           # open the Memory Palace

# .hebbsignore (same syntax as .gitignore)
templates/
drafts/*.md
archive/
*.template.md

hebbs start                           # gRPC :6380, HTTP :6381
hebbs remember "hello world"          # store a memory (uses server via --endpoint)
hebbs recall "hello"                  # recall it

from hebbs import HebbsClient

client = HebbsClient("localhost:6380")

await client.remember(
    content="Prospect mentioned competitor contract expires March 15",
    importance=0.95,
    entity_id="acme",
    context={"stage": "discovery", "signal": "urgency"},
)

# Four recall strategies
history = await client.recall(cue="acme engagement", strategy="temporal", entity_id="acme")
responses = await client.recall(cue="we built this in-house", strategy="similarity")
causes = await client.recall(cue="deal lost after pricing", strategy="causal")
patterns = await client.recall(cue="healthcare compliance objection", strategy="analogical")

# Consolidate and query insights
result = await client.reflect()
insights = await client.insights(entity_id="acme", max_results=10)

import { HebbsClient } from '@hebbs/sdk';

const client = new HebbsClient('localhost:6380', { apiKey: process.env.HEBBS_API_KEY });
await client.connect();

await client.remember({
    content: 'Prospect mentioned competitor contract expires March 15',
    importance: 0.95,
    entityId: 'acme',
    context: { stage: 'discovery', signal: 'urgency' },
});

// Four recall strategies
const history = await client.recall({ cue: 'acme engagement', strategies: ['temporal'], entityId: 'acme' });
const causes = await client.recall({ cue: 'deal lost after pricing', strategies: ['causal'] });
const patterns = await client.recall({ cue: 'healthcare compliance objection', strategies: ['analogical'] });

// Consolidate and query insights
const result = await client.reflect();
const insights = await client.insights({ entityId: 'acme', maxResults: 10 });

pip install hebbs[demo]
hebbs-demo interactive --config gemini-vertex --verbosity verbose

cd hebbs-typescript/demo && npm install
npx tsx src/index.ts interactive --mock-llm

hebbs --tenant acme-corp remember "Q2 forecast looks strong" --entity-id project-alpha

client = HebbsClient("localhost:6380", tenant_id="acme-corp")

const client = new HebbsClient('localhost:6380', { tenantId: 'acme-corp' });

──────────────────────────────────────────────────────────
                     Client SDKs
             Python  |  TypeScript  |  Rust
──────────────────────────────────────────────────────────
               Agent Skills (SKILL.md)
            Claude Code  |  OpenClaw
──────────────────────────────────────────────────────────
                gRPC  |  HTTP/REST
──────────────────────────────────────────────────────────

                  Core Engine (Rust)

  +------------+ +------------+ +------------------+
  |  Remember   | |   Recall   | | Reflect Pipeline |
  |  Engine     | |   Engine   | | (background)     |
  |             | |            | |                  |
  | - encode    | | - prime    | | - cluster (Rust) |
  | - score     | | - query    | | - propose (LLM)  |
  | - index     | | - subscribe| | - validate (LLM) |
  | - decay     | | - merge    | | - store insights |
  +------+------+ +------+-----+ +--------+---------+
         |               |                |
  +------+---------------+----------------+-----------+
  |              Index Layer                          |
  |   Temporal (B-tree) | Vector (HNSW) | Graph       |
  +----------------------+----------------------------+
                         |
  +----------------------+----------------------------+
  |         Storage Engine (RocksDB)                  |
  |         Column Families per index type            |
  +---------------------------------------------------+

  +-----------------------+  +------------------------+
  | Embedding Engine      |  | LLM Provider Interface |
  | (ONNX Runtime,        |  | (Anthropic, OpenAI,    |
  |  built-in default)    |  |  Ollama, pluggable)    |
  +-----------------------+  +------------------------+

hebbs start                                       # gRPC :6380, HTTP :6381
HEBBS_AUTH_ENABLED=true hebbs start                # with API key authentication

from hebbs import HEBBS

e = HEBBS.open("./agent-memory")  # No separate process
e.remember(...)