Skip to content

Memory

Memory is injected behind the MemoryManager seam. The ContextBuilder retrieves in-scope items and injects them into the planner context (emitting memory_retrieved / memory_injected). Core ships three reference managers.

InMemoryMemory — term-overlap relevance

Retrieval ranks in-scope items by how many query terms appear in the content (a question surfaces a fact that shares words), not substring containment. preference items are always floored in so standing instructions surface. Ties break by item score, then recency. For tests and local examples.

EmbeddingMemory — semantic recall

Ranks in-scope items by cosine similarity between the query embedding and each item's content embedding. The embedder is injected — no model is bundled, so it runs in any runtime. A text-less query falls back to score/recency.

import { EmbeddingMemory, type Embedder } from "@agentic-kernel/core";

const embed: Embedder = async (text) => myEmbeddingModel(text); // number[]
const memory = new EmbeddingMemory({ embed });
// new AgentEngine({ planner, policy, memory, ... })

from agentic_kernel.core import EmbeddingMemory

async def embed(text: str) -> list[float]:
    return await my_embedding_model(text)

memory = EmbeddingMemory(embed=embed)
# engine = create_agent_engine(planner=..., policy=..., memory=memory)

Context-driven recall

Since 0.6.1, the recall query follows the run instead of repeating the opening task string. Each round, ContextBuilder composes the retrieval query from:

  1. the task input (opening intent),
  2. the latest user message (if it differs from the task input), and
  3. snippets of recent successful observation results (where the run actually went),

capped at 2000 characters (observation snippets at 200 each). A run that starts "fix the homepage layout" and drifts to "wire the staging DB" now recalls staging-DB memory mid-run instead of only homepage memory.

On the first round (no messages or observations yet) the query is exactly task.input, so turn-1 behavior is unchanged. This is internal query construction — no API or schema change; hosts get it automatically.

Scopes

Memory is tenant-namespaced: a scope is tenant/[workspace/][agent]. A seeded item's scope must be within the agent's scope prefix to be retrievable — e.g. for tenant acme, agent researcher, an item with scope acme/researcher matches.

Production & governance

For durable storage use memory-postgres (or any MemoryManager). Long-term memory has a governance lifecycle — proposeWrite / supersede / tombstone / explain / export — and SessionMemoryExtractor / a memory governor can extract attributable session summaries automatically on completion.