Memory Modes: EXTRACTED vs VERBATIM vs HYBRID

Choose how HippoDid processes and stores memories for each character.

Table of contents

Overview
EXTRACTED (default)
VERBATIM
HYBRID
Comparison table
Setting the memory mode
Choosing the right mode
Storage cost comparison
Next steps

Overview

Every HippoDid character has a memory mode that controls how add_memory processes incoming content. The mode is set when the character is created and can be changed at any time.

Mode	Processing	Best for
EXTRACTED	AI categorizes and extracts structured facts	Most use cases (default)
VERBATIM	Stores exact text as-is, no AI processing	Compliance, legal, exact quotes
HYBRID	AI extraction + verbatim archive	Maximum fidelity when budget allows

EXTRACTED (default)

When you call add_memory with content like:

“The client prefers to be contacted by email, not phone. They use Slack for internal comms and have a renewal date of March 2027.”

EXTRACTED mode runs the content through HippoDid’s AI pipeline, which:

Splits the text into individual facts
Assigns each fact a category (e.g., preferences, events)
Scores salience (0.0 to 1.0)
Detects and resolves conflicts with existing memories
Generates vector embeddings for semantic search

Result: three separate memories stored:

Category	Content	Salience
preferences	Client prefers email contact, not phone	0.75
preferences	Client uses Slack for internal communication	0.60
events	Client renewal date: March 2027	0.85

When to use: Most of the time. EXTRACTED gives the best search results, the cleanest memory organization, and automatic deduplication.

VERBATIM

VERBATIM mode stores the exact text you provide with no AI processing. The content is saved as a single memory entry with the category you specify (or uncategorized if you do not).

Using the same input:

“The client prefers to be contacted by email, not phone. They use Slack for internal comms and have a renewal date of March 2027.”

Result: one memory stored:

Category	Content	Salience
uncategorized	The client prefers to be contacted by email, not phone. They use Slack for internal comms and have a renewal date of March 2027.	0.50

When to use:

Compliance and legal: you need an exact record of what was said, with no AI paraphrasing
Audit trails: regulators require the original text, not a summary
Exact quotes: customer quotes, verbatim feedback, specific instructions
Cost control: no AI processing means no AI operation costs

Trade-offs:

No automatic categorization or fact splitting
No conflict detection or deduplication
Search relies on keyword overlap rather than semantic understanding
You get back exactly what you put in, nothing more

HYBRID

HYBRID mode does both: it runs the full EXTRACTED pipeline and also archives the original verbatim text. You get the structured, searchable facts plus an unmodified copy.

Using the same input, HYBRID produces:

Category	Content	Salience	Type
preferences	Client prefers email contact, not phone	0.75	extracted
preferences	Client uses Slack for internal communication	0.60	extracted
events	Client renewal date: March 2027	0.85	extracted
verbatim_archive	The client prefers to be contacted by email, not phone. They use Slack for internal comms and have a renewal date of March 2027.	0.50	verbatim

When to use:

Regulated industries: you need AI-powered search and retrieval but also need to keep exact originals for audit
Session transcripts: store the structured facts for the agent and the raw transcript for compliance
High-value interactions: customer success calls, legal consultations, medical notes

Trade-offs:

Uses the most storage (roughly 2x compared to EXTRACTED alone)
Costs one AI operation per add_memory call (same as EXTRACTED)
The verbatim archive is searchable but not as well-organized as the extracted facts

Comparison table

	EXTRACTED	VERBATIM	HYBRID
AI processing	Yes	No	Yes
Fact splitting	Yes	No	Yes
Auto-categorization	Yes	No	Yes
Conflict detection	Yes	No	Yes
Semantic search quality	Best	Basic	Best
Exact original preserved	No	Yes	Yes
Storage per input	Low	Lowest	Highest
AI operation cost	1 op	0 ops	1 op
Compliance-ready	Partial	Full	Full

Setting the memory mode

At character creation

curl -X POST https://api.hippodid.com/v1/characters \
  -H "Authorization: Bearer hd_key_..." \
  -H "Content-Type: application/json" \
  -d '{
    "name": "Legal Client - Smith",
    "categoryPreset": "standard",
    "memoryMode": "VERBATIM"
  }'

Updating an existing character

curl -X PATCH https://api.hippodid.com/v1/characters/CHARACTER_ID \
  -H "Authorization: Bearer hd_key_..." \
  -H "Content-Type: application/json" \
  -d '{
    "memoryMode": "HYBRID"
  }'

Python SDK

from hippodid import HippoDid

client = HippoDid(api_key="hd_key_...")

# Create the character first
character = client.create_character(
    name="Legal Client - Smith",
)

# Then set memory mode (must be set after creation, not at create time)
client.set_memory_mode(character.id, "VERBATIM")

# Switch to HYBRID later
client.update_character(
    character_id=character.id,
    memory_mode="HYBRID",
)

Via MCP tools

If you are using HippoDid through Claude Code, Cursor, or another MCP client:

Use hippodid to set the memory mode of "Legal Client - Smith" to HYBRID

The MCP set_memory_mode tool accepts EXTRACTED, VERBATIM, or HYBRID.

Choosing the right mode

Start with EXTRACTED (the default). It gives the best search results and keeps memory organized automatically. Switch only when you have a specific reason.

Use VERBATIM when:

You are in a regulated industry (healthcare, legal, finance) and auditors need exact records
You want zero AI processing cost
You are storing structured data that should not be paraphrased (JSON, code snippets, exact quotes)

Use HYBRID when:

You need both: great search results from extraction and exact originals for compliance
You are logging high-value customer interactions where both the structured facts and the raw transcript matter
Budget is not a constraint and you want maximum fidelity

Storage cost comparison

Assuming 1,000 memory inputs per month, each averaging 200 words:

Mode	Memories stored	AI ops	Relative storage
EXTRACTED	~3,000 (after fact splitting)	1,000	1x
VERBATIM	1,000 (one per input)	0	0.7x
HYBRID	~4,000 (extracted + verbatim)	1,000	1.7x

Exact numbers depend on the density of facts in your inputs. Inputs with many distinct facts produce more extracted memories.

Next steps

Assembly Strategies — how different strategies format memories for prompts
Batch Onboarding — set memory mode for thousands of characters at once via templates
Quick Start — create your first character
API Reference — full endpoint documentation