Skip to content

README.md

Latest commit

 

History

History
169 lines (126 loc) · 7.37 KB

README.md

File metadata and controls

169 lines (126 loc) · 7.37 KB
Raw

hotmemory

Agent memory as tables on Hotdata.

A memory record is a row in a managed table. The row carries the text, a scope, tags, source references, and the time span in which the fact was true. The table is an ordinary Hotdata table, so an agent can join its memory to its own data in one SQL query. None of the memory systems that we surveyed stores memory as typed columns in the same engine as the data of the consumer.

Status: version 0.0.0, not published. The storage contract exists in Python with two drivers. MemoryStore runs in process memory. HotdataStore keeps records in one Hotdata managed database. The memory contract, Memory, runs over either driver. docs/internal/roadmap.md lists the phases.

The library has two layers:

  • A storage contract. Put, get, list, search, and delete records in a namespace. Records are immutable. A new put under the same key creates a new revision.
  • A memory contract. Remember facts, recall them inside a context budget, supersede a fact, forget by id or by horizon, and load the profile of a subject. Extraction from raw text is optional, and it takes an extractor callable that the caller supplies. load cuts a Markdown document into chunks, stores each chunk as an episode, and extracts facts that name their episode as a source. recall and profile return facts, profiles, and procedures, never episodes.

The library runs in the process of the consumer and calls the Hotdata API with the API key of the consumer. There is no hotmemory server. The first consumer is an incident investigator built on Hotdata. The library is not specific to it. A plain Python agent, a LangGraph agent, or any process with a Hotdata API key can use it.

Try the storage contract

MemoryStore needs an embedder for a search with query text. An embedder is a callable that turns a list of texts into a list of vectors. This example uses a toy embedder that counts two words:

from hotmemory import Filter, MemoryStore


def embed(texts):
    return [[text.count("disk") + 0.1, text.count("cpu") + 0.1] for text in texts]


store = MemoryStore(embedder=embed)
store.put(("team", "alerts"), "disk", kind="fact", content="The disk fills at night.")
store.put(("team", "alerts"), "disk", kind="fact", content="The disk fills at noon.")
store.put(("team", "alerts"), "cpu", kind="fact", content="The cpu spikes after a deploy.")

print(store.get(("team", "alerts"), "disk").id)  # team/alerts/disk@2
print([r.revision for r in store.history(("team", "alerts"), "disk")])  # [1, 2]
hits = store.search("disk", [("team",)], Filter(kind="fact"), k=1)
print(hits[0].record.content)  # The disk fills at noon.

Try the memory contract

Memory gives an agent the memory operations over any store. remember derives each key from the subject and the content, so a retried call writes nothing new. recall returns the records and one block of text, with one line for each record:

from hotmemory import Fact, Memory, MemoryStore


def embed(texts):
    return [[text.count("disk") + 0.1, text.count("cpu") + 0.1] for text in texts]


memory = Memory(MemoryStore(embedder=embed))
scope = ("team", "alerts")
memory.remember(
    [Fact(kind="fact", subject="disk", content="The disk fills at night.", sources=("chat-1",))],
    scope,
)
records, block = memory.recall("disk", [scope])
print(block)  # - The disk fills at night. [sources: chat-1] [valid: unknown to now]

load takes a document name, the Markdown text, a scope, and an extractor. The extractor gets each chunk, with the first heading and the heading path of the chunk at its top:

from hotmemory import Fact, Memory, MemoryStore


def embed(texts):
    return [[text.count("disk") + 0.1, text.count("cpu") + 0.1] for text in texts]


def last_line(text, observed_at, current):
    return [Fact(kind="fact", subject="disk", content=text.splitlines()[-1])]


memory = Memory(MemoryStore(embedder=embed))
text = "# Disk incident\n\n## Root cause\n\nThe disk fills at night.\n"
episodes, facts = memory.load("disk-incident", text, ("team", "alerts"), last_line)
print(episodes)  # ['team/alerts/disk-incident-0001@1']
episode = memory.store.get(("team", "alerts"), "disk-incident-0001")
print(episode.content.splitlines()[0])  # # Disk incident

skills/hotmemory/SKILL.md gives an agent the same operations as commands.

Store memory in Hotdata

HotdataStore needs the hotdata extra, and an embedder for every write and search. The openai extra ships OpenAIEmbedder, which reads OPENAI_API_KEY. Any callable that turns a list of texts into a list of vectors also works. provision opens the database with the name, or creates it. It reads the connection from HOTDATA_API_KEY, HOTDATA_WORKSPACE, and HOTDATA_API_URL:

from hotmemory.hotdata import HotdataStore
from hotmemory.openai import OpenAIEmbedder

embedder = OpenAIEmbedder()
store = HotdataStore.provision(
    "agent-memory", embedder=embedder, model=embedder.model, dimensions=1536
)
store.put(("team", "alerts"), "disk", kind="fact", content="The disk fills at night.")
print([hit.record.id for hit in store.search("disk", [("team",)])])

One process writes to a database. docs/contracts.md gives the tables, the retrieval query, and the rules of the driver. docs/local.md tells you how to run it against a local engine.

Benchmark

hotmemory ran LongMemEval-S, a public benchmark of 500 questions over long chat histories, on 2026-10-10. docs/benchmarks.md gives every number with its configuration.

Retrieval needs no model except the embedder. The runner stored each session of a question as an episode in HotdataStore and searched with the question. The 419 scored questions had these results with the vector ranking. For 98.3% of them, an evidence session was in the top 5. For 99.3% of them, every evidence session was in the top 10.

The end-to-end score ran on a sample of 54 questions. The reader and the judge were gpt-4o-2024-08-06.

What the reader got Correct
Every session of the question, about 114,000 tokens 39 of 54 (72.2%)
The 10 sessions that the search ranked first 45 of 54 (83.3%)
The block of recall, after capture with a gpt-4o-mini extractor 19 of 54 (35.2%)

The memory contract, with this extractor, scores below the full history. In 19 of its 35 misses, the extractor did not store the answer. With 54 questions, each score can be about 10 points off. In the cloud, a search took 361 ms at p50 with the vector ranking, and 998 ms with the fused ranking.

Documents

  • docs/contracts.md: the record, the storage operations, the memory operations, and the platform facts behind them.
  • docs/guarantees.md: each behavior that a consumer can rely on, its state, and its proof.
  • docs/local.md: how to run the library against a local RuntimeDB container.
  • docs/benchmarks.md: the numbers of hotmemory on LongMemEval-S, each with its configuration. bench/README.md tells you how to run it.
  • CONTRIBUTING.md: the one command that checks a change, and the rules of the test suite.
  • CHANGELOG.md: each change to a public surface.
  • docs/internal/: the design brief, the survey, the roadmap, and the plan for the current phase.